Artificial intelligence system and method for processing data-sets using robust rolling frameworks

Abstract

An approach for reducing and extracting meaningful signals from large data-sets is disclosed. The approach comprises generating a data-set having a first dimension representing a rolling window in a time series, a second dimension representing number of financial instruments, a third dimension representing number of observations through a predetermined time period, and a fourth dimension representing number of features. The approach further comprises clustering the data set; setting cluster size and rolling window length; and applying K-Means algorithm to the data-set to establish a plurality of centroids. The approach also comprises for each of the plurality of centroids, selecting closest data point in a current data set, wherein each centroid has at least one data point. The approach further comprises iteratively applying the K-Means algorithm according to the selected data points until the corresponding centroids converge; and determining scores for the centroids to assess separation of the clusters, wherein the scores are utilized to determine performance of the financial instruments. Moreover, the approach comprises identifying a plurality of regimes corresponding to statistical models, and computing a plurality of scores for the statistical models. The approach further comprises comparing the statistical models using marginal likelihood, and selecting one of the statistical models based on the plurality of scores. Furthermore, the approach comprises selecting one of a plurality of financial instruments based on the selected statistical model.

Description

Attorney Docket No.: P10375PC00 PatentARTIFICIAL INTELLIGENCE SYSTEM AND METHOD FORPROCESSING DATA-SETSUSING ROBUST ROLLING FRAMEWORKSRELATED APPLICATIONS

[0001] This application claims the benefit of U.S. Provisional Patent Application No.63 / 554,546, titled “Artificial Intelligence System and Method for Processing Data-Sets UsingRobust Rolling Frameworks,” filed February 16, 2024, the entire disclosure of which is herebyincorporated by reference herein.BACKGROUND

[0002] Despite the current set of tools, data analysts in all industries face the daunting task ofanalyzing an overwhelming barrage of information to determine meaningful, impactful data. Oneparticular sector involves the financial industry, such as wealth management. The selection of vastnumber of profitable assets and investment portfolios require allocation of enormous resources inpersonnel and technology. Existing systems that wealth managers rely on for financial analysishave not kept based with development in data processing technology. The challenge is magnifiedwhen the data (e.g., news, social media, etc.) significantly affects market conditions in real-time.SOME EXAMPLE EMBODIMENTS

[0003] Therefore, there is a need for an approach that applies artificial intelligence (AI), e.g.,machine learning and deep learning, for data analysis to support decision making.

[0004] According to one embodiment, a method comprises generating a data-set having a firstdimension representing a rolling window in a time series, a second dimension representing numberof financial instruments, a third dimension representing number of observations through apredetermined time period, and a fourth dimension representing number of features. The methodfurther comprises clustering the data set; setting cluster size and rolling window length; andAttorney Docket No.: P10375PC00 Patentapplying K-Means algorithm to the data-set to establish a plurality of centroids. The method alsocomprises for each of the plurality of centroids, selecting closest data point in a current data set,wherein each centroid has at least one data point. The method further comprises iterativelyapplying the K-Means algorithm according to the selected data points until the correspondingcentroids converge; and determining scores for the centroids to assess separation of the clusters,wherein the scores are utilized to determine performance of the financial instruments.

[0005] According to one embodiment, a method comprises identifying a plurality of regimescorresponding to statistical models, and computing a plurality of scores for the statistical models,wherein each of the plurality of scores measures similarity of a data point to a corresponding clusteragainst dissimilarity of another cluster. The method further comprises comparing the statisticalmodels using marginal likelihood, and selecting, for machine learning, one of the statistical modelsbased on the plurality of scores. Furthermore, the method comprises selecting one of a pluralityof financial instruments based on the selected statistical model.

[0006] In addition, for various example embodiments of the invention, the following isapplicable: a method comprising facilitating a processing of and / or processing (1) data and / or (2)information and / or (3) at least one signal, the (1) data and / or (2) information and / or (3) at least onesignal based, at least in part, on (or derived at least in part from) any one or any combination ofmethods (or processes) disclosed in this application as relevant to any embodiment of theinvention.

[0007] For various example embodiments of the invention, the following is also applicable:the data-set includes financial data relating to funds, wherein the items of the second dimensioncorrespond to the funds.

[0008] For various example embodiments of the invention, the following is also applicable:the processing of the data-set further comprises generating cosine similarity scores for the data-set; and altering eigenvectors at each increment to account for sign change and eigenvectordirections according to the cosine similarity scores.

[0009] For various example embodiments of the invention, the following is also applicable: amethod comprising facilitating creating and / or facilitating modifying (1) at least one device userinterface element and / or (2) at least one device user interface functionality, the (1) at least onedevice user interface element and / or (2) at least one device user interface functionality based, atAttorney Docket No.: P10375PC00 Patentleast in part, on data and / or information resulting from one or any combination of methods orprocesses disclosed in this application as relevant to any embodiment of the invention, and / or atleast one signal resulting from one or any combination of methods (or processes) disclosed in thisapplication as relevant to any embodiment of the invention.

[0010] In various example embodiments, the methods (or processes) can be accomplished onthe service provider side or on the mobile device side or in any shared way between the serviceprovider and mobile device with actions being performed on both sides.

[0011] For various example embodiments, the following is applicable: An apparatuscomprising means for performing a method of any of the claims.

[0012] Still other aspects, features, and advantages of the invention are readily apparent fromthe following detailed description, simply by illustrating a number of particular embodiments andimplementations, including the best mode contemplated for carrying out the invention. Theinvention is also capable of other and different embodiments, and its several details can bemodified in various obvious respects, all without departing from the spirit and scope of theinvention. Accordingly, the drawings and description are to be regarded as illustrative in nature,and not as restrictive.BRIEF DESCRIPTION OF THE DRAWINGS

[0013] The embodiments of the invention are illustrated by way of example, and not by wayof limitation, in the figures of the accompanying drawings:

[0014] FIG. 1 is a diagram of an AI-based decision support platform, according to oneembodiment;

[0015] FIG. 2 is a diagram of the components of the AI-based decision support platform ofFIG. 1, according to one embodiment;

[0016] FIG. 3 is a diagram of the AI-based decision support platform of FIG. 1 interactingwith various data sources, according to one embodiment;

[0017] FIG.4A is a flowchart of a process for reducing and extracting meaningful signals fromlarge data-sets, according to one embodiment;Attorney Docket No.: P10375PC00 Patent

[0018] FIG. 4B is a flowchart of a process for assessing performance data by the AI-baseddecision support platform of FIG. 1, according to one embodiment;

[0019] FIG. 4C is a flowchart of a process for executing robust rolling K-Means (R2K-Means)by the AI-based decision support platform of FIG. 1, according to one embodiment;

[0020] FIG. 4D is a flowchart of a process for executing robust rolling regime detection (R2-RD) by the AI-based decision support platform of FIG. 1, according to one embodiment;

[0021]

[0022] FIGs. 5A-5H are diagrams of a graphical user interface (GUI) relating to datamanagement processes performed by the AI-based decision support platform of FIG. 1, accordingto one embodiment;

[0023] FIGs. 6A-6D are diagrams of a GUI relating to asset allocation processes performed bythe AI-based decision support platform of FIG. 1, according to one embodiment;

[0024] FIGs. 7A-7C are diagrams of a GUI relating to security selection processes performedby the AI-based decision support platform of FIG. 1, according to one embodiment;

[0025] FIG. 8 is a diagram of a neural network that can be implemented by the AI-baseddecision support platform of FIG. 1, according to one embodiment;

[0026] FIG. 9 is a diagram of hardware that can be used to implement various exampleembodiments;

[0027] FIG. 10 is a diagram of a chip set that can be used to implement various exampleembodiments; and

[0028] FIG. 11 is a diagram of a mobile terminal (e.g., handset) that can be used to implementvarious example embodiments.DESCRIPTION OF SOME EMBODIMENTS

[0029] Examples of a method, apparatus, and computer program for reducing and extractingmeaningful signals from large data-sets to utilize as training data are disclosed. In the followingdescription, for the purposes of explanation, numerous specific details are set forth in order toAttorney Docket No.: P10375PC00 Patentprovide a thorough understanding of the embodiments of the invention. It is apparent, however,to one skilled in the art that the embodiments of the invention may be practiced without thesespecific details or with an equivalent arrangement. In other instances, well-known structures anddevices are shown in block diagram form in order to avoid unnecessarily obscuring theembodiments of the invention.

[0030] FIG. 1 is a diagram of AI-based decision support platform, according to oneembodiment. Clustering techniques are commonly used in the financial industry to group data orassets for isolating behaviors (based on performance measures, holdings, or alternative data). K-Means is a well-known clustering algorithm designed to partition a data set into k clusters. Timeseries datasets from the financial markets are constantly being updated when new informationbecomes available. Updating clusters simply by re-running standard K-Means on the new data willproduce unstable and chaotic results due to the effects of random initialization sensitivity andcluster mislabeling.

[0031] Moreover, financial markets are inherently challenging to navigate due to theirpronounced non-stationarity and high dimensionality. The concept of regimes offers one way ofbringing some order to these complexities. This perspective assumes the existence of a smallnumber of distinct regimes, each marked by similar market behaviors. For instance, some fundsor strategies may excel during inflationary times, yet falter in deflationary periods. Suchvariability underscores the importance for allocators to discern how to manage capital at risk.Where understanding the defining attributes of these regimes, as well as the dynamics of transitionsbetween them, becomes paramount. Such insights are critical across all facets of the InvestmentLife Cycle, including security selection, asset allocation, portfolio construction, asset planning,and risk management. Implementations can also benefit other financial processes, including credit,risk, policies, etc.

[0032] To address the noted drawbacks of conventional systems and approaches to processingthe vast amount of data for proper selection of investments, a system 100 of FIG. 1 includes anAI-based decision support platform 101 that introduces the capability to assess and to measurereal-time impact of an asset / security or portfolio. The R2K-Means is to create a stable andupdatable framework in which K-Means centroids can adapt to new time series data.Attorney Docket No.: P10375PC00 Patent

[0033] The motivation of R2K-Means method stems from the financial industry’s need ofupdateable and stable clustering techniques to dynamically model securities throughout time.Additionally, groupings of financial assets may form irregular shapes and patterns in the financialmarkets requiring non-linear clustering techniques for meaningful results. The platform 101, underone scenario, can provide an advisor the capability to build a portfolio by investing in a cluster andupdating the advisor’s portfolio when new data becomes available. Naive K-Means will randomlyforce the investor to liquidate and reinvest their entire portfolio purely due to random initialization.Furthermore, this updated cluster may exist in a new location and represent a completely differentbehavior in asset performance.

[0034] Furthermore, the nonstationary and high-dimensional nature of financial markets posessignificant challenges for navigation. Temporally stable regime classification offers a perspectiveto manage these challenges. Consequently, a Robust Rolling Regime Detection (R2-RD)framework is provided by platform 101; such framework adaptively retrains with streaming dataand employs temporal ensemble, label assignment, and threshold policies to address temporalinstability resulting from nonstationarity, model mismatches, etc.

[0035] Additionally, the platform 101 can supplement and / or integrate with a third partysystem 103, which may be a conventional financial analysis system that is utilized by financialadvisors. The platform 101 can parse through numerous data-sets, models, viewpoints, visuals,etc. concurrently to continually assess historic and predicted performance within each aspect ofthe asset life cycle, subject to defined or suggested objective functions. By way of example, theplatform 101 can supplement the financial advisor’s experience-based, traditional portfoliodecisions with Artificial Intelligence-powered visualizations for assessing a Security or Portfolio(A), evaluating the predicted path profile of A, comparing A with B (and transition of A to B), andmeasuring real-time impact on A. With respect to the third party system 103, the platform 101integrates within the technology infrastructure of the system 103 to execute user-defined oroptimized objective functions and can also run the processes on existing portfolios.

[0036] Among other functions and features, the platform 101 can harness research, data,programming and advanced AI methodologies to address daunting development possibilitiescentered around the following: removing the drudgery of complex tasks to drive efficiency;Attorney Docket No.: P10375PC00 Patentproviding visuals that are interpretable to facilitate more creative work; and determining meaningresearch data.

[0037] As shown in FIG. 1, the system 100 also comprises user equipment (UE) 105a-105n(collectively referred to as UE 105) that may include or be associated with applications 107a-107n(collectively referred to as applications 107). In one embodiment, the UE 105 has connectivity tothe AI-based decision support platform 101 via the communication network 109. Under certainscenarios, the AI-based decision support platform 101 performs one or more functions associatedwith data analysis in conjunction with the UEs 105a-105n. By way of example, the UE 105 is anytype of mobile terminal, fixed terminal, or portable terminal including a mobile handset, station,unit, device, multimedia computer, multimedia tablet, Internet node, communicator, desktopcomputer, laptop computer, notebook computer, netbook computer, tablet computer, personalcommunication system (PCS) device, personal navigation device, personal digital assistants(PDAs), audio / video player, digital camera / camcorder, positioning device, television receiver,radio broadcast receiver, electronic book device, game device, a smartphone, a smartwatch, smarteyewear, or any combination thereof, including the accessories and peripherals of these devices,or any combination thereof. It is also contemplated that the UE 105 can support any type ofinterface to the user (such as “wearable” circuitry, etc.). In one embodiment, the UE 105 mayinclude Global Positioning System (GPS) receivers to obtain geographic coordinates fromsatellites (not shown) for determining current location and time associated with the UE 105; suchGPS information can be utilized to geo-tag images captured by UE sensors (not shown). The UE105 is capable of supporting a graphical user interface (GUI) that provides the GUI of FIGs. 5-7.

[0038] The AI-based decision support platform 101 operates in conjunction with one or moreapplications resident on an UE 105. By way of example, the applications 107 may be any type ofapplication that is executable at UE 105, such as content provisioning services, camera / imagingapplication, media player applications, social networking applications, calendar applications, andthe like. In one embodiment, the applications 107 may assist in conveying sensor information viathe communication network 109. In another embodiment, one of the applications 107 at the UE105 may act as a client for the AI-based decision support platform 101 and perform one or morefunctions associated with the functions of the platform 113 by interacting with the platform 113over the communication network 109.Attorney Docket No.: P10375PC00 Patent

[0039] One or more data sources 109a-109n are accessible via the network 109 by the platform101. The data sources 109a-109n can include any content relevant to making financial decisionsrelating to securities, such as market data (e.g., historical or real-time market information), news,financial holdings, and / or alternative data. The retrieved data can reside within database 111 ofthe AI-based decision support platform 101. It is contemplated that database 111 can beimplemented as a cloud storage system.

[0040] Further details of the capabilities of the AI-based decision support platform 101 isprovided in FIGs. 3-7.

[0041] The communication network 109 of system 100 includes one or more networks suchas a data network, a wireless network, a telephony network, or any combination thereof. It iscontemplated that the data network may be any local area network (LAN), metropolitan areanetwork (MAN), wide area network (WAN), a public data network (e.g., the Internet), short-rangewireless network, or any other suitable packet-switched network, such as a commercially owned,proprietary packet-switched network, e.g., a proprietary cable or fiber-optic network, and the like,or any combination thereof. In addition, the wireless network may be, for example, a cellularnetwork and may employ various technologies including 5G (5th Generation), 4G, 3G, 2G, LongTerm Evolution (LTE), enhanced data rates for global evolution (EDGE), general packet radioservice (GPRS), global system for mobile communications (GSM), Internet protocol multimediasubsystem (IMS), universal mobile telecommunications system (UMTS), etc., as well as any othersuitable wireless medium, e.g., worldwide interoperability for microwave access (WiMAX), codedivision multiple access (CDMA), wideband code division multiple access (WCDMA), wirelessfidelity (Wi-Fi), wireless LAN (WLAN), Bluetooth®, Internet Protocol (IP) data casting, satellite,mobile ad-hoc network (MANET), and the like, or any combination thereof.

[0042] In one embodiment, the AI-based decision support platform 101 may be a platformwith multiple interconnected components. The AI-based decision support platform 101 mayinclude multiple servers, intelligent networking devices, computing devices, components andcorresponding software for providing real-time data analysis. In addition, it is noted that the AI-based decision support platform 101 may be integrated or separated from services platform. Also,certain functionalities of the system 101 may reside within the UE 105 (e.g., as part of theapplications 107).Attorney Docket No.: P10375PC00 Patent

[0043] Moreover, the platform 101 can interface with various services systems (not shown),such as notification services, content (e.g., audio, video, images, etc.) provisioning services,application services, storage services, contextual information determination services, socialnetworking services, location-based services, information-based services, etc.

[0044] By way of example, UE 105, the AI-based decision support platform 101, the thirdparty system 103 with each other and other components of the communication network 109 usingwell known, new or still developing protocols (e.g., IoT standards and protocols). In this context,a protocol includes a set of rules defining how the network nodes within the communicationnetwork 109 interact with each other based on information sent over the communication links.The protocols are effective at different layers of operation within each node, from generating andreceiving physical signals of various types, to selecting a link for transferring those signals, to theformat of information indicated by those signals, to identifying which software applicationexecuting on a computer system sends or receives the information. The conceptually differentlayers of protocols for exchanging information over a network are described in the Open SystemsInterconnection (OSI) Reference Model.

[0045] Communications between the network nodes are typically effected by exchangingdiscrete packets of data. Each packet typically comprises (1) header information associated witha particular protocol, and (2) payload information that follows the header information and containsinformation that may be processed independently of that particular protocol. In some protocols,the packet includes (3) trailer information following the payload and indicating the end of thepayload information. The header includes information such as the source of the packet, itsdestination, the length of the payload, and other properties used by the protocol. Often, the datain the payload for the particular protocol includes a header and payload for a different protocolassociated with a different, higher layer of the OSI Reference Model. The header for a particularprotocol typically indicates a type for the next protocol contained in its payload. The higher layerprotocol is said to be encapsulated in the lower layer protocol. The headers included in a packettraversing multiple heterogeneous networks, such as the Internet, typically include a physical(layer 1) header, a data-link (layer 2) header, an internetwork (layer 3) header and a transport (layer4) header, and various application (layer 5, layer 6 and layer 7) headers as defined by the OSIReference Model.Attorney Docket No.: P10375PC00 Patent

[0046] FIG. 2 is a diagram of the components of the AI-based decision support platform ofFIG. 1, according to one embodiment. By way of example, the platform 101 includes one or morecomponents for analyzing data to determine selection and / or performance of a security / asset orportfolio. It is contemplated that the functions of these components may be combined in one ormore components or performed by other components of equivalent functionality. In thisembodiment, the AI-based decision support platform 101 includes the following modules: an assetallocation module 201, a portfolio construction module 203, a risk management module 205, afinancial planning module 207, a portfolio management module 209, a security selection module211, and an artificial intelligence engine 213.

[0100] The asset allocation module 201 generates a multitude of dynamic allocation / weightsbased on historic or predicted objective functions (e.g., minimal return of x, max MDD of y, etc.)based on market indices and / or other market data. Additionally, the module 201 can incorporatemacro-economic and social data, thereby providing “explainability” (e.g., indicia) of howallocations are made (e.g., weighting of measures, macro condition, alternative, etc.). The module201 has the capability to change weights or the complete selection criteria, and perform “what if”analyses; this is conducted for selection, tracking or comparative purposes. Further, the module201 provides assessment of historic and ongoing tracking error of allocations. Optionally, themodule 201 can employ clusters (as opposed to traditional asset class definitions). For example,security clusters can have behavioral, holding or alternative data similarity (for the duration of theassessed period). In effect, the module 201 can satisfy the objective function(s) given evolvingmarket conditions over a particular time horizon.

[0101] According to one embodiment, multi-stage processing is utilized: behavioral, assettype, and firm level. Behavioral clustering identifies the commonalities of funds based on fundbehavior rather than the holding data. Performance measures are passed into the behavioralprocessing stage. A deep temporal cluster model finds the behavioral commonalities of thesecurities (e.g., funds). With Asset Type processing, funds are clustered (grouped) based on theirholding data. For instance, the model takes as inputs holding data, in which a hierarchicalclustering model finds the optimal number of clusters. A K-means model finds outliers and mergethem together. In the Firm Level stage, clustering utilizes alternative data to find commonalitiesof at the firm level. By way of example, first, static data is applied in a Tree-based model to groupAttorney Docket No.: P10375PC00 Patentfunds based on funds stable characteristics. Secondly, monthly data is applied to a K-means modelto find the non-stable commonalities.

[0102] As part of indexing, cluster labels are generated to provide a clear outlook on the assetholding and economic characteristics for each cluster. According to one embodiment, labels areassigned based on the user’s objective and can be used as inputs for other modules substitutingtraditional measures and data categories. The clusters, according to one embodiment, throughStatistical Measure, Economic Meaning, and Stability. Regarding Statistical Measure, each clustergets validated by different metrics to assure its mathematical correctness. With respect toEconomic Meaning, the economic characteristics of each fund in a cluster are compared to assurethe validity of the labels. As for Stability, the process is repeated, by way of example, for differenttime horizons to determine the stability of the cluster. Clusters’ labels are thus relatively stable toprovide value to other modules.

[0047] Among other functions, the portfolio construction module 203 generates dynamicportfolio compositions based on historic or predicted objective function (e.g., min return of x, maxMDD of y, etc.). The module 203 utilizes the outputs of the asset allocation module 201 and thesecurity selection module 211. As with the asset allocation module 201, the module 203 providesthe capability to demonstrate how the composition was made (e.g., relationships, macro condition,alternative, etc.), as well as configure the selection criteria and modify weighting parameters. Themodule 203 also supports assessment of historic and ongoing tracking error of portfolios; errorcorrection and penalizing functions are employed for reducing tracking error. Moreover, themodule 203 provides an option to create or optimize portfolios based on clusters, objectivefunction, factors, minus indices, or other criteria.

[0048] With respect to the risk management module 205, this module 205 focuses on theactionable points on a historic, predictive and real-time basis. For instance, the following real-time assessment are supported: portfolio and security sensitivity to factors (e.g., value, trend, etc.);security / portfolio / holdings exposure and give indicators for loss or hedge thresholds and actiontriggers; opinions within a given text across news, social media etc. as it relates to the fund,portfolio, holdings, factor, etc. Additionally, the module 205 can provide sentiment analysis onAttorney Docket No.: P10375PC00 Patentunstructured text into structured data using, for example, Natural Language Processing (NLP).Further, the module 205 utilizes a mathematically stable risk proxy for portfolio comparisons.

[0049] The financial planning module 207 provides accurate forecast of cash flows of assetsunder various scenarios. For example, the module 207 can generate combined cash flow profilesusing a common set of assumptions (e.g., macro, performance measure, etc.). Additionally, themodule 207 can conduct stress and scenario analysis on the combined cash flow profiles, factors,etc. The module 207 utilize multi-step and machine learning (ML) illiquid forecasting models forfinancial planning.

[0050] The portfolio management module 209 enables management functions relating to aparticular asset, holding, and / or portfolio based on machine learning / deep learning (ML / DL)framework with continuous additions of data, measures and visuals per user requests. Thefunctions of module 209 can be a service that is cloud based, standalone or embedded with anexisting system (e.g., third party system 103).

[0051] The platform 101 provides evaluation, via the security selection module 211, of thebest-performing funds / securities within an asset class, filtered pool or cluster for a given objectivefunction. Such objective function can be based on queries / filters that also combine multiple themes(including pre / post filtering). In one embodiment, the security selection module 211 performsdynamic ranking of securities based on historic or predicted objective functions (e.g., min returnof x, max MDD of y, etc.). Ranking allows the selection of a smaller number of funds / securitiesout of a large number of candidates based on the objective function. For example, among the 500funds, the top 10 best-performing funds can be selected by the module 211. For ranking, ML-based ranking can be utilized solely or in combination with empirical ranking. Empirical rankinguses weighted average of the performance measures, either historical or predicted, to do simpleranking and selection. ML-based ranking can include interpretable ML models and online learningalgorithms. Interpretable models can include models such as Decision Trees or RuleFit to findbest ranking methodologies. The features can be either historical or predicted performancemeasures. Online learning algorithms can include weighting expert, which dynamically selectsreliable models and signals for ranking and selection.Attorney Docket No.: P10375PC00 Patent

[0052] The module 211 can supplement traditional section criteria of performance measureswith macro-economic, alternative data and holding data. The module 211 supports a capability toexplain to the user why the security was selected (e.g., measures, macro condition, alternative,etc.). The user can specify user-defined performance measures, change weights, or modify thecomplete selection criteria. As with the other modules, the security selection module 211 cansupport selection, tracking or comparative purposes, and provide assessment of historic andongoing tracking error of the selection.

[0053] The artificial intelligence engine 213 interact with one or more of the various modules201-211 to support the functions of the platform 101. The AI engine 213, according to oneembodiment, utilizes a Robust Rolling Transformer Text Classifier (R2-T2C) model, which usesnatural language processing (NLP) techniques to create numerical representations (embeddings)of a corpus of text. Each document will be split into chunks of text prior to embedding and pre-processed. With an embedded corpus, R2-T2C generates cluster embeddings with a list ofkeywords with the option of processing the word list through a prompt. The key word list can bedesigned to summarize various targets / categories of words commonly used in that subject. Next,similarity scores, classifications, and insight can then be drawn from the embeddings of both thecorpus and the classification clusters.

[0054] R2-T2C utilizes the transformer encoder mode BERT and creates a numericalrepresentation of text sequences (tokens) using both the token type and position as input. Positionalencodings help enrich the models understanding of meaning and context. The addition of a promptas a preprocessing step for cluster generation improves the embedding locations and subsequentclassification performance of R2-T2C. R2-T2C improves on this framework by improvingcomputational efficiency, creating new visuals and tests, and implementing a prompt to boostclassification performance. Additionally, R2-T2C adds the ability to compare documents throughvisuals and the Hellinger distance measure. This process is applied to policy documents / guidelinesto how the corpus of documents should be written. A performance test of generating clustersthrough the corpus metadata allows the parameters of R2-T2C to be optimized.

[0055] Table 1 depicts an exemplary R2-T2C Framework:Attorney Docket No.: P10375PC00 Patent1. Load the text corpus created by the pdf reader / web scraper process;2. Choose a sentence length for the document to be split into;3. Split the documents into sentences while applying text preprocessing functionsto clean;and prepare the text for embedding;4. Choose an embedding model (BERT recommended) and embed the corpus;5. Preprocess and embed clusters by applying text formatting and embeddingsimilar to the text corpus;6. Select a similarity measure (cosine similarity, Euclidean distance, etc) tocompare each sentence in the corpus to each cluster;7. Classify all sentences in the corpus and determine classification of document bycombining sentence results onto an overall score;8. Classify all sentences in the corpus and determine classification of document bycombining sentence results onto an overall score;9. Analyze and visualize the results of the classification; and10. Test R2-T2C using the corpus metadata on multiple parameter sets for cross-validation and model optimization.Table 1

[0056] By way of example, the AI engine 213 can execute the neural network of FIG. 8.

[0057] The above presented modules and components of the AI-based decision supportplatform 101 can be implemented in hardware, firmware, software, or a combination thereof.Though depicted as a separate entity in FIG. 1, it is contemplated that the AI-based decisionsupport platform 101 may be implemented for direct operation by respective UE 105. As such,the AI-based decision support platform 101 may generate direct signal inputs by way of theoperating system of the UE 105 for interacting with the applications 107. In another embodiment,one or more of the modules of FIG. 2 and processes of FIGs. 3, 4, and 7 may be implemented foroperation by respective UEs, the AI-based decision support platform 101, or combination thereof.Still further, the AI-based decision support platform 101 may be integrated for direct operationAttorney Docket No.: P10375PC00 Patentwith services 115, such as in the form of a widget or applet, in accordance with an informationand / or subscriber sharing arrangement. The various executions presented herein contemplate anyand all arrangements and models.

[0058] FIG. 3 is a diagram of the AI-based decision support platform of FIG. 1 interactingwith various data sources, according to one embodiment. The platform 101 architecturally iscreated in a modular fashion with extensive artificial intelligence-powered data manipulation andanalytical capabilities. In one embodiment, the platform 101 has a flexible architecture and iscloud-based. With respect to securities, such as mutual funds and exchange traded funds (ETFs),the platform 101 can collect robust and cleaned data-set. For instance, the platform 101 cleansand processes data from various sources 301 (e.g., market, alternative, client, social, etc.).Additionally, the platform 101 provides the capability to incorporate / append other third-party / unique data-sets and advisor holdings. In terms of processing, the platform 101 is modularbased on user (e.g., financial advisor) needs and application programming interfaces (APIs) 303for user ecosystem connectivity. The platform 101 has the capability to select data-sets, filtersand features for assessment.

[0059] Advantageously, the platform 101 provides a seamless, building block framework withcontinually added depth and breadth. The platform 101 performs an optimized internal processingthat mixes-matches AI techniques / models within the modules best suited for a given objectivefunction. The platform 101 measures tracking error, assess “explainability” of results and abilityto custom train. Module / version release / accretive analysis is readily provided as the platform 101via modules 307 operates from Security to Portfolio to Risk Management (and vice versa)presented in visually friendly, non-technical way.

[0060] The platform 101 covers critical portfolio decision facets. For instance, the platform101 monitors real time market signals - jumps, trends, sentiment analysis for sectors, industry.Traditionally, training is extremely time-consuming to perform manually, thus, the platform 101,in one embodiment, employs auto hyperparameter search or neural architecture search methods.For example, basic hyperparameter search methods include: Grid search, Random search, andBayesian optimization.Attorney Docket No.: P10375PC00 Patent

[0061] By way of example, various different options for selecting regimes for model trainingand / or simulations can be utilized by platform 101: (1) auto selected model based on historicaldata (below classification compared with S&P500™ daily returns); (2) machine learning modelsthat predict the future regime (based on factors, economic indicators, other); and / or user views onthe regimes across other modules (e.g., security or portfolio performance).

[0062] The platform 101 also performs Security / manager selection based on objectivefunctions - historic or predictive, as well as Asset allocation based on market indices or securities(or security proxies). Optimized or behavioral finance-based custom portfolio construction can beperformed using, for example, traditional methods, clusters or factors. Portfolios can be replicatedusing other securities (e.g., replicating mutual fund portfolios using ETFs and / or otheroptimizations). The platform 101 can assess custom or uploaded portfolios or transitions withvaried holdings. Risk management can be overlaid on market / sector / portfolio / security. Further,the platform 101 can perform: Stress testing / Scenario analysis; Factor mapping / sensitivity acrossall aspects; NLP connect – news / factors; NLP sentiment – news / factors; and / or Risk Watch –alert / hedging / stop loss.

[0063] The platform 101 provides comprehensive forecasts for financial planning includingincorporating illiquid assets. In this regard, the platform 101 provides: performance / results acrossvarious wealth management modules; amplifies existing advisor methods / tools; continualimplementation of new research within modules; methods / metrics and publishing research; anentirely AI-based with no legacy technical issues; pool of advanced AI-developers and financialmarket specialists; an eco-system - engaging renowned AI and finance experts as advisors.

[0064] FIG.4A is a flowchart of a process for reducing and extracting meaningful signals fromlarge data-sets, according to one embodiment. In one embodiment, the AI-based decision supportplatform 101 performs the process 420 and is implemented in, for instance, a chip set including aprocessor and a memory as shown in FIG. 16. Per step 401, the platform 101 generates a data-sethaving multiple dimensions including a first dimension representing a rolling window in a timeseries, a second dimension representing number of funds, a third dimension representing numberof observations through a predetermined time period, and a fourth dimension representing numberof features. The data-set is processed, as in step 423, by recursively determining principalAttorney Docket No.: P10375PC00 Patentcomponents of a reduced data-set from a portion of the data-set, wherein the principal componentsare determined at each increment in the rolling window. In step 426, the platform 101 generatestraining data based on the processed data-set, and outputs the training data to a learning model (perstep 427) – e.g., AI engine 213.

[0065] FIG. 4B is a flowchart of a process for assessing performance data by the AI-baseddecision support platform of FIG. 1, according to one embodiment. In one embodiment, the AI-based decision support platform 101 performs the process 400 and is implemented in, for instance,a chip set including a processor and a memory as shown in FIG. 16. As shown, per step 401, theplatform 101 collects data from one or more data sources related to an asset or a portfolio. Theplatform 101 also analyzes historic performance data for the asset or the portfolio, per step 403.In step 405, predictive performance data is generated for the asset or the portfolio. The platform101 continually generates assessment data based on the historic performance data and thepredictive performance using an artificial intelligence engine, as in step 407. The generatedassessment data is output via a graphical user interface of UE 105a, for example (step 409).

[0066] FIG. 4C is a flowchart of a process for executing robust rolling K-Means (R2K-Means)by the AI-based decision support platform of FIG. 1, according to one embodiment. Process 430involves, per step 431, generating a data-set (similarly to the process of FIG. 4A); the data-set hasa first dimension representing a rolling window in a time series, a second dimension representingnumber of financial instruments (e.g., funds), a third dimension representing number ofobservations through a predetermined time period, and a fourth dimension representing number offeatures. As in step 433, the process 430 clusters the data set; whereby cluster size and rollingwindow length are set. A K-Means algorithm is then applied to the data-set to establish a pluralityof centroids (per step 435). For each of the plurality of centroids, the closest data point in a currentdata set is selected, wherein each centroid has at least one data point. Process 430 then iterativelyapplies, as in step 439, the K-Means algorithm according to the selected data points until thecorresponding centroids converge. Per step 441, scores for the centroids are determined to assessseparation of the clusters, wherein the scores are utilized to determine performance of the financialinstruments. Details of process 430 are further provided as follows.Attorney Docket No.: P10375PC00 Patent

[0067] The time series clustering typically focuses on datasets arranged with time T as itsfeature space (columns) and groups the rows F (funds or individual time series) according to somemeasure of similarity. In terms of notation, time series models cluster thetogroup all F into k clusters. K-Means is a spatial clustering algorithm and struggles to cluster Xwith time as its feature space since it cannot capture the shift and scaling characteristics of timeseries. Common in the financial industry, assets have a feature space D relating to variousperformance measures and other attributes. R2K-Means is designed for time series data structuredas , where K-Means clustering can be iteratively run on for all t = 1,…,Tat each time increment. This structure is moreto the spatialnature of K-Means and allows for additional increments of time to be added to the data set withoutchanging previous results. Centroid-based clustering models operating on this data structure areupdatable and will produce a time series of centroids. R2K-Means utilizes a rolling window ofcentroids influenced by both space and path to create nonlinear decision boundaries forclassification at each t = 2,…, T time increment.

[0068] The standard K-Means algorithm (below) begins by randomly selecting k data pointsfrom the data set as the initial centroids of the model. K-means will then group all data points byproximity to the nearest centroid provided a distance metric, the most common of which beingEuclidean distance. Next, the centroids will be updated as the average of all data points in thatcluster. This iterative process will continue until all centroids have converged. Each data point isthen provided the label of the centroid it is closest to.clustering. Improvements to the random initialization step of K-Means, such as K++, have beendeveloped advantageously to encourage cluster separation and boost computational performance.Attorney Docket No.: P10375PC00 PatentIt is important to note that nearly all centroid initialization methods rely on some form of randomselection. This is a challenge when attempting to update clusters when new time series databecomes available. K-Means is traditionally designed to cluster datasets X as arrays of twodimensions indexed by (F, D), with F denoting funds (data points) and D as its features. K-Meansgroups all funds in F into a specified k different clusters. In this study of time series analysis, thetime dimension T is added turning X into a tensor indexed by (F, D, T). In financial applications,special considerations must be made when navigating the dimension of time. New data from themarket is constantly incrementing the size of T, resulting in new groupings of funds F in futuretime increments.

[0070] Consider the naive K-Means approach to updating the groupings of funds as new timebecomes available. Starting at t = 1, the standard K-Means algorithm is applied to Xt=1 where thefunds are grouped accordingly. Then at t = 2,…, T the same process is applied, producing agrouping of funds at each point of time. These time indexed groupings can model the dynamicrelationship between funds across time. Funds should change clusters if its performance in thefeature space location drifts closer to a different centroid. In practice, the groupings of funds willvary wildly across time for three primary reasons: (1) Random Initialization Sensitivity: K-Meansrandomly selects different initial starting locations for the centroids, resulting in differentconverged centroid locations; (2) Cluster Mislabeling: The cluster labels are arbitrarily assignedeach K-Means run, potentially resulting in, for example, Cluster 1 being labeled as Cluster 2 in thenext time period and vice versa; and (3) Changes in Data: The underlying data changes throughouttime. The objective of the R2K-Means framework is to minimize the effects of items 1 and 2 sothe clustering results will purely model the changes in data across time.

[0071] Computer-based implementations of K-means algorithm assign arbitrary labels foreach cluster after the algorithm has converged. In a time series setting, there arise problems inrelating which cluster is which between differing time increments. The random initialization stepof K-Means typically names the centroid labels in the order in which the centroid was assigned. Ifdata point xi is randomly selected first then it will be given the label of Cluster 1 (0 for computers),and so on. Standard K-Means random initialization will provide each data point an equalprobability of being selected first. It is note that a dataset with two data points, x1 and x2, and K-Means can be used to cluster this dataset into k = 2 clusters. The centroids will be located at eachAttorney Docket No.: P10375PC00 Patentof the two data points. The centroid labels may change each time K-Means is used to cluster thedataset since data point x1 has a 50% chance of being randomly selected first. Even though thecentroid locations do not change, the labels for them will and produce unstable clustering labelsacross time.

[0072] A subtle but important aspect of K-Means is that there should be an ordering fromwhich the centroids are initialized. Centroid locations should be unique for all clusters duringrandom initialization. If a dataset contains duplicate data points, then randomly selecting points asstarting centroids may result in duplicate centroids which will break K-Means. R2K-Means utilizesthe cluster labels directly from the previous time increment. Accordingly, there is no label relatingor estimation required since the cluster labels are deterministically brought to the current state ofiteration.

[0073] R2K-Means leverages a robust rolling technique applied to the K-Means algorithmfrom time t = 2 and onward. The first period will simply use the standard K-Means algorithm (orany of its variants) as desired. The focal alteration to K-Means is to replace the randominitialization step by selecting the data point closest to each of the centroids from the previousperiod. The ordering in which each centroid selects its closest data point is determined by a numberof data points it held from the last period (ascending order). Once each centroid has been assignedto its closest data point, then the rest of the standard K-means algorithm will run, resulting in newlocations for each centroid. This eliminates the need for cluster label estimation since the algorithmdirectly pulls the centroid labels from the previous period. The implementation of R2K-Meansbelow is a direct modification of the standard K-Means algorithm where the converged centroidsin t = 2, …, T only reference the information from the previous time period. The labeling step willuse centroids from the last W time increments, where W is the length of the rolling window.One of the advantages of R2K-Means is that from the second period onward, the only source ofrandomness is from the dataset itself. If the data does not change, then the clustering results willnot change when the algorithm is updated. This property introduces a degree of path dependencyto the otherwise spatial nature of K-Means since the centroids react only to changes in the data.The R2K-Means algorithm is provided as follows:Attorney Docket No.: P10375PC00 Patent

[0074] The R2K-Means design process was guided by two general rules for ensuringsuccessful algorithm convergence at each time step: (1) Shared Centroid Location: The centroids(means) can never share the same location in space of another centroid; and (2) Centroid Sparsity:Each centroid must be closest to at least one data point in each iteration. If a centroid has no closestdata points, it leads to a division by zero error when recalculating its position, as the calculationaverages the positions of these points.

[0075] K-Means satisfies both rules in the random initialization process (step 3 in the abovealgorithm). By randomly selecting a data point as the starting location of the centroid, each centroidis ensured to have at least one data point. When K-means is randomly selecting data points duringrandom initialization, it must validate that all data points are unique. If two data points are selectedin the same location, then the model will break as it violates rule 1. K-Means does not randomlyselect all centroids at once, instead it must choose the starting centroid locations in a randomordering during initialization.

[0076] R2K-Means uses the centroids from the previous period as starting locations, thenassigns each centroid to the data point closest to it. Since standard K-Means at t = 1 establishes thecentroids do not share location, using these centroids at t = 2 will also adhere to rule 1 and so on.Attorney Docket No.: P10375PC00 PatentR2K-Means will then assign each centroid to the data point closest to it assuming that point hasnot already been claimed by another centroid (by ascending order). By design, this step makescertain that all centroids will contain at least one data point. Afterwards, the rest of the conventionalK-Means algorithm will commence until each centroid converges.

[0077] It is contemplated that additional blueprints for an updating K-Means model can beutilized, including a model which would linearly interpolate between time increments and runmultiple K-Means until both rules were satisfied. While the interpolation model can be utilized,issues can arise if new data would abandon or disappear from a centroid. With no data to interpolateto, a centroid would be left with zero data points and violate rule 2. As such, R2K-Means providesdeterministic stability, lower computational complexity, and robustness.

[0078] While a single K-Means centroid forms a linear decision boundary, a combination ofcentroids sharing the same label can form complex nonlinear shapes. For nonlinear classification,R2K-Means uses a rolling window of centroids from the last W time periods to classify data in thecurrent time step. This adjustment adds a degree of path dependency to K-Means and can fitdatasets with more exotic shapes in the financial market. There are many considerations to makewhen selecting the size of the rolling window. If nonlinearity is not desired, simply setting the W= 1 will turn R2K-Means into a linear classification algorithm. Important to note, at t = 1 eachcentroid will be linear regardless of window length. The larger the rolling window, the higherchange that rolling window of centroids will form nonlinear shapes. In the scope of computationalcomplexity, each time increment will require the distances between each data point must becalculated for the k x W centroids. Centroids will only change due to changes in data, so datawhich shifts frequently after normalization is more likely to create nonlinear decision boundaries.The longer the W, the more chance that centroids of differing labels and time will overlap due topath dependency. This results in disconnected decision boundaries for each cluster. It isrecommended to use a rolling window length between, e.g., 6-12, to best achieve nonlinear andconnected clusters.

[0079] Regarding the ordering of centroid selection, after the centroids from the previous arebrought to the current, they will then be assigned to the data point closest to it. Since each centroidmust exist in a unique location, there must be an ordering in which the centroids are assigned. AnAttorney Docket No.: P10375PC00 Patentoptimal ordering method to assign this is to count the number of data points from the previousperiod and rank in ascending order. This will give the centroid with the fewest data points the firstpriority in selecting its closest data point. Clusters with the fewest data points are the mostvulnerable to having their close data points taken before their turn in the ordering. Therefore, thebest way to reduce variance and promote model stability is to sort centroids in ascending orderbased on data count (see steps 8, 19 of R2K-Means Algorithm 2).

[0080] K-Means clustering has numerous financial applications outside of grouping assetsby performance. For example, K-Means centroids are commonly used as input for regime detectionalgorithms such as Gaussian Mixture Models and Hidden Markov Models. The stable and reliableproperties of R2K-Means centroids can be passed on to regime models granting the ability toupdate regime predictions when new information becomes available. A rolling window basedregime model using R2K-Means has the ability to incorporate both long term (centroids) and shortterm (rolling window) aspects of data into a robust framework.

[0081] In certain embodiments, Robust Rolling Regime Detection (R2-RD) is a HiddenMarkov Model utilizing the centroids of R2K-Means (W = 1) as a stable starting location in whichthe Gaussian distributions are fitted. The regime results of Naive K-Means are highly unstable andfrequently change due to random initialization and cluster mislabeling, rendering the regimeprobabilities unsuitable for financial application. The regime classifications of R2-RD and R2K-Means are much more chronologically stable and well separated. R2K-Means produces a timeseries of centroids primarily influenced by space with a small amount of path dependency. Sincecentroids are themselves a time series, distance metrics more applicable to time series can beutilized including Dynamic Time Warping. It is contemplated that alterations to R2K-Means canbe made to make the centroids more path driven, allowing for a longer rolling windows to producepath-related clusters across time. Additional processing steps could be added to ensure the rollingwindow of centroids do not overlap to mitigate unconnected decision boundaries. R2K-Meanscould be adapted to a variable number of k clusters throughout time to better fit the data at each tincrement.

[0082] FIG. 4D is a flowchart of a process for executing Robust Rolling Regime detection(R2-RD) by the AI-based decision support platform of FIG. 1, according to one embodiment. AsAttorney Docket No.: P10375PC00 Patentnoted, the nonstationary and high-dimensional nature of financial markets poses significantchallenges for navigation. Temporally stable regime classification offers a perspective to managethese challenges. A Robust Rolling Regime Detection (R2-RD) framework is provided by theplatform 101 to adaptively retrain with streaming data and employs temporal ensemble, labelassignment, and threshold policies to address temporal instability resulting from nonstationarity,model mismatches, etc. Further, the R2-RD framework’s data-driven model selection procedurechooses the model that best describes the data from the wide variety of latent variable models.

[0083] Process 450 identifies, according to one embodiment, a plurality of regimescorresponding to statistical models, as in step 451. In step 453, scores (e.g., Silhouette scores) arecomputed for the statistical models. The statistical models are compared using marginal likelihood(per step 455). Process 450 then selects, for machine learning, one of the statistical models basedon the scores, per step 457. Each of the scores measures similarity of a data point to acorresponding cluster against dissimilarity of another cluster. Furthermore, one of a plurality offinancial instruments (which in certain embodiments include securities) is selected based on theselected statistical model (as in step 459).

[0084] Details of process 430 are further detailed as follows. Financial regimes, by theirnature, are not directly observable. Hence, latent variable models become an ideal choice for theirdiscovery. Among these, the family of Hidden Markov Model (HMM) stands out as particularlyeffective. At its core, a standard (“vanilla”) HMM presumes a finite number of hidden states(which, in our study, are termed as regimes) whose transition dynamics are Markovian, while theobservations are conditionally independent given the hidden states. Variations of HMM relax theassumptions from the vanilla version, such as Markovian regimes and conditional independenceof the observables, leading to models like the Hidden Semi-Markov Model (HSMM), Auto-Regressive HMM (ARHMM, also known as the Markov Switching Model), etc. The platform 101aims to delineate the unique features of each regime with these models and derive a time seriesthat indicates the trajectory of realized regimes based on given observations.

[0085] When a model is applied to data, certain assumptions are inherently relied upon.However, because no assumption is perfectly valid in real-world scenarios, it is important toconsider a range of models, thereby enabling selection of the one that most appropriately fits theAttorney Docket No.: P10375PC00 Patentspecific situation. However, selecting the best model is not straightforward, especially forunsupervised tasks, like regime detection. The process described herein leverages a combinationof classic statistical concepts, the marginal likelihood (or its approximation), and a geometricmeasure, the Silhouette score, to choose the model that best describes the data. The marginallikelihood of a model measures the probability of observing the data conditioning on the model,which helps with selection among different families of models, like the different variations ofHMM mentioned above. The Silhouette score then determines the optimal number of regimesgiven a family of models to maximize inter-cluster separation and intra-cluster similarity.

[0086] Fitting a model once to all the time series data ignores the internal dynamics of howthe market evolves across time. It is recognized that there is the need for rolling- or expanding-window retraining (or fine-tuning) with financial time series data to gain insights about temporalinstability. Temporal instability refers to the unstable data generation process in financial markets,which summarizes concepts such as nonstationarity and non-Markovian. Temporal instability iseven more challenging with optimization errors from solving a nonconvex model. A local optimumfound from training the model in a first period could drastically differ from the one found fromretraining the model in a second period. Such inconsistency is a result of both temporal instabilityand optimization error. To address this inconsistency issue, R2-RD uses temporal ensembletechniques, which smooth the models fitted consecutively in order to alleviate the inconsistencyissue from optimization error.

[0087] Another challenge R2-RD addresses is labeling. Like most unsupervised learningmodels, a regime detection model, such as a standard HMM, generates several unlabeled clusters(regimes). The order of regimes in which the algorithm returns depends on the initialization andoptimization processes, which are mostly stochastic by construction. Labels can be assigned toregimes according to such orderings. In that case, a random assignment can be performed, whichcan be troublesome during retraining, as there will be mismatches between the previously foundregimes and the newly refitted ones. For example, regime 1 from the previously fitted HMMdescribes a high gross domestic product (GDP) environment, while regime 1 from the newly fittedHMM could correspond to a low GDP environment. A label assignment model is introduced toresolve this problem without affecting the internal optimization process of the regime models. Thelabel assignment model is an integer programming that minimizes the total costs of assigning theAttorney Docket No.: P10375PC00 Patentlabels from previously defined regimes to the newly found ones, where the costs are measured bystatistical distances.

[0088] Moreover, the platform 101 aims to determine the appropriate number of regimes.Although the model selection framework partially addresses this issue by comparing the Silhouettescore, this approach primarily applies to the initial training period. For instance, suppose the firsttraining session selects a HMM with 4 regimes; however, in subsequent periods, the score mightsuggest a HMM with, e.g., only three regimes. One objective is to maintain historical consistency:once a regime is established, it should remain present and not vanish. Therefore, a threshold policyis introduced; the policy relies on optimal label assignment costs to determine the number ofregimes in ongoing recursive training sessions following the initial fit. Such policy helps indetermining whether a new regime emerges in the market based on recent observations.

[0089] By way of illustration, the efficacy of R2-RD is demonstrated using two distinctdatasets: macroeconomic and futures markets. In the macroeconomic context, the processidentifies a few unique regimes. Further analysis of regime characteristics, through rollingretraining, reveals shifting distributions over time for each regime. This shift underscores thenecessity for dynamic decision models in financial markets. A similar pattern can be observed inthe futures markets. Moreover, such analysis indicates a pronounced distinction in mutual fundperformances across different macroeconomic regimes, highlighting the substantial potential ofregime-aware asset management strategies.

[0090] According to certain embodiments, the platform 101 provides a framework thatutilizes the family of HMM for online regime detection, emphasizing the need for smart modelselection and introducing a procedure based on statistical and geometric measures. R2-RD alsoadopts temporal ensemble, label assignment, and a threshold policy that alleviates the temporalinstability issue and identifies the emergence of new regimes. In addition, the platform 101provides strong empirical evidence of nonstationary market dynamics that can only be found by arecursive framework.

[0091] Under R2-RD, a sample of multi-dimensional time series data is considered.The observation in each period Xt is a D-dimensional column vector. It isthat there existsan unobserved discrete-time stochastic process with a finite state space Z. ConsideringAttorney Docket No.: P10375PC00 Patentthe case of macroeconomic regimes, the D-dimensional time series data is the collection ofmacroeconomic indicators of, e.g., Gross Domestic Product (GDP), unemployment rate,Consumer Price Index (CPI), etc. In examining GDP and CPI, one investment clock (e.g., MerrillLynch™) defines the state space as: Z = {Reflation, Recovery, Overheat, Stagflation}. Usingstatistical models like HMM, more than just GDP and CPI are considered, whereby regimeinformation can be extracted from a much larger feature space.

[0092] The vanilla HMM with Gaussian observation model provides a good benchmark latentvariable model for regime detection. It assumes that the hidden state Zt follows a Markov chain,and each observation Xt is conditionally independent given the hidden state. Gaussian observationmodel refers to the fact that Xt follows a D-dimensional Gaussian distribution, whose parametersdepend on the hidden state Zt. Variations of HMM are obtained by either relaxing the Markovianassumption or designing a more complicated observation model. For example, Gaussian mixtureHMM uses the Gaussian mixture as the observation model. Autoregressive HMM relaxes theconditional independence assumption for observation and assumes that the parameters of Xt’sdistribution not only depend on Zt but also Xt−1. Each model makes certain assumptions about theunderlying data generation process. Thus, a model selection procedure is introduced for selectingthe model that best describes the data statistically and also best separates the regimesgeometrically.

[0093] HMM and its variations are single-stage models in that it is not designed for retrainingor fine-tuning with streaming data. Thus, the process introduces temporal ensemble, labelassignment model, and a threshold policy to ensure determination of a stable trajectory of regimedynamics. Altogether, this framework is defined as the Robust Rolling Regime Detection (R2-RD).

[0094] Model selection remains a crucial aspect in machine learning, as it enables criticalevaluation and justification of the underlying assumptions of various candidate models. Therefore,the platform 101 emphasizes selecting models through data-driven methods, employing widelyrecognized techniques such as cross-validation. However, in many applications in unsupervisedlearning, social science, etc., cross-validation provides little help because the ground truth (e.g.,label) is not known. Consequently, a key statistical concept is incorporated for model selection:Attorney Docket No.: P10375PC00 Patentmarginal likelihood. This metric inherently assists with selecting the model that most closely alignswith the data. The underlying principles are described as follows,

[0095] Firstly, marginal likelihood is defined: consider a set of models (hypothesis) (M1, . . ., MK), parameterized by (θ1, ... , θK) respectively. Given data , the marginal likelihoodof model Mi is defined as:where Pr(Y|Mi, θi) is the likelihood (or outcome model) and π(θi|Mi) is the prior distribution ofparameters θi. This value quantifies the probability of observing the data set Y given model Mi.Next, it is shown why marginal likelihood can be used for model selection. Applying the Bayesformula, the following results:where Pr(Mi) is the prior distribution for model i, and it is typically defined to be discrete uniformso that Pr(Mi) = Pr(Mj) for all i and j. Pr(Mi|Y) is the posterior distribution and the value of interest,which describes the probability that Mi is the model generating the data given the observed data.

[0096] Finally, in order to select the model that best describes the data, the posterior odds arecalculated:The right-hand side is also denoted the Bayes factor. It is noted that model comparison reduces theproblem of comparing marginal likelihood. That is, the higher the marginal likelihood, the betterthe model can describe the observed data. When the Bayes factor between two models, Mi and Mjis close to 1, indicating that the data does not significantly differentiate between the two models,alternative metrics can be considered for model selection -- the Silhouette score can be utilized. InAttorney Docket No.: P10375PC00 Patentgeneral, the calculation of the marginal likelihood is complicated. A prominent approximation ofthe marginal likelihood is the Bayesian Information Criterion (BIC) score. Assuming unitinformation prior and using the central limit theorem for posterior distribution, the log of marginallikelihood reduces to the Schwarz criterion, minus twice of which is the BIC score:BICi = −2 log Pr(Y|Mi, θi) + d logNwhere d is the number of parameters of model Mi. Hence, a model with a larger marginal log-likelihood is approximately equivalent to a smaller BIC score. Then, if the precise marginallikelihood is unavailable, we could select models with a small BIC score.

[0097] However, marginal likelihood could have unsatisfying performance if the priordistribution is not chosen appropriately. Such issues can lead to ambiguous selection amongmodels that are close to each other. BIC approximation has the same selection ambiguity issue.Thus, marginal likelihood-based selection is applied among model families. For example, HMMwith Gaussian observation model, HMM with Gaussian mixture observation model, HSMM, andARHMM, etc. For comparison within a family of models, i.e., hyper-parameter selection,measures with geometric meanings are utilized: the Silhouette score. The Silhouette scoremeasures the similarity of a data point to its own cluster against the dissimilarity of it to otherclusters. It provides a more intuitive comparison for models with different numbers of clusters(regimes). Thus, the model selection procedure is summarized as the following process, per Table2: 1. Compare model families using the marginal likelihood (or BIC approximation) and choosethe one with the largest likelihood (or smallest BIC score)2. Within the selected model family, choose the model with the best geometric separation (thehighest Silhouette score)Table 2

[0098] The described model selection procedure is designed with the purpose of describingthe data in the most statistically and geometrically meaningful way. If the focus is to improve theAttorney Docket No.: P10375PC00 Patentperformance of some downstream tasks like security selection and portfolio construction, then itis better to customize a selection process that serves the best interest of the downstream tasks.

[0099] Many latent variable models have nonconvex objective functions, solving which couldlead to potentially large optimization errors (sub-optimality). Although the local optima might beclose to the global one in the case of vanilla HMM due to the Baum-Welch algorithm, suchguarantees are not ensured for other more complicated variations. A direct consequence ofoptimization error is that when solving a model with rolling window data, the sequence of locallyoptimal solutions could be far away from each other, leading to a highly volatile trajectory ofregimes. Thus, to improve the robustness of the framework and find a stable trajectory of regimes,the impact of optimization error needs to be mitigated.

[0100] Essentially, the sequence of models needs to smoothed from rolling window fitting.Smoothing in the Hilbert space is complicated, consequently a temporal ensemble is introduced -- a set of heuristics for smoothing statistical models fitted across time. One of the techniques istemporal initialization, where one approach, from a fitted model, is used to initialize the subsequentmodel. An HMM with the Gaussian observation model is used as an example. It is assumed thatthe model is fitted on a sub-sequence of data with 2 regimes; denoted the regimes foundas and . Movingto time t2, new data is acquired, and thedata . Instead of randomor initializing with K-means, which is a commonHMM, the Expectation-Maximization algorithm isinitialized at the previously found regimes . If the model is being solved forthe first time, a simple optimization by randomly choosing multipleinitializations and selecting the one thatAccording to one embodiment,this scheme is only applied to the first training because such process can be time-consuming.Moreover, the consistency issue does not exist when there are no fitted models to compare with.

[0101] Another useful temporal ensemble method is averaging model parameters across time.This is helpful, particularly when dealing with forecasting problems. That is, when the transitionmatrix from HMM type model is used to forecast future regime distributions, (exponentially)averaging the transition matrix from multiple training alleviates the impact of temporal instabilityand optimization errors.Attorney Docket No.: P10375PC00 Patent

[0102] With respect to label management, a generalization of the previous example isconsidered. When fitted to, a set of regimes denoted as I is obtained. When fitted, a new set of regimes denoted as J is obtained. If the label of an old regime i ∈ I is assigned to anew regime j ∈ J, it will incur an assignment cost cij , measured by the statistical distance betweenthe two regimes. In the case where the old and new regimes are modeled as two Gaussiandistributions, an intuitive choice for the distance measure would be a weighted difference betweenthe first two moments:α is a hyperparameterthe first and second moments.denotes the Frobenius norm. Because the first two moments are sufficient statistics for adistribution, the above expression is general enough for quantifying the distance betweenGaussian distributions.

[0103] However, in the case of more complicated distributions, other statistical distances areutilized. According to certain embodiments, distances from the information geometry literature(e.g., Kullback-Leibler distance) or optimal transportation costs (e.g., Wasserstein distance) aregood choices for the assignment cost.

[0104] The decision variables of the label assignment problem is the assignment matrix withentries xij ∈ {0, 1} indicating whether the label of old regime i is assigned to new regime j. Theassignment problem can then be formulated as follows:Attorney Docket No.: P10375PC00 Patent

[0105] The first constraint specifies that, at most, one label is assigned to a new regime. If|J| > |I| and a new regime does not get assigned, then it is defined as an emerging new regime. Thesecond constraint specifies that each label of the old regime is assigned to exactly one new regime.

[0106] Although the normalization constant |I| in the objective does not affect the solution, itis maintained because it makes the objective value an average assignment cost that is comparableunder different numbers of regimes. This is important for the threshold policy.

[0107] In certain embodiments, the threshold policy for determining emergence of newregimes is detailed as follows.

[0108] One drawback of single-stage models like HMM is a constant state space. It isrecognized that there could be new regimes emerging in the market, which cannot be modeled bytraditional HMM. Thus, the optimal objective values from the label assignment problem is usedas a criterion for determining if there is a new regime appearing. This problem is not treated as ageneral hyper-parameter tuning because the method employed herein explicitly models anemerging new regime as a geometrically distant distribution from the existing ones. The conceptof geometrically distant is captured by the label assignment cost, which measures the total distancefrom the old regimes to the new regimes. In addition, only either one regime or no regime emergingis considered. The reason is that with granular enough data (typically monthly frequency formacroeconomic indicators and minute-level frequency for futures data), it is improbable to havemore than one new regime emerging.

[0109] Specifically, given |I| regimes found in , the following models on dataare solved: one with the number of regimes equal the other with |I| + 1. labelassignment problem is solved for both models – the optimal objective c0 and c1,respectively. If the assignment cost of adding one more regime is smaller and theassignment cost of maintaining the same number of regimes is higher than a certain threshold, then it is considered a new regime in the market. The intuition isthe distancebetween the old and new regimes is too large and adding one more regime to the new ones candecrease the distance, it is considered an emerging regime situation.

[0110] The above described R2-RD process can be applied to macroeconomic regimedetection, whereby R2-D2 alleviates the temporal instability issues and uncovers interestingAttorney Docket No.: P10375PC00 Patentpatterns in the macroeconomic environment. By way of example, macroeconomic indicators fromthe Federal Reserve Economic Data (FRED) can be utilized, whereby the dataset covers a largevariety of measures, including the GDP of various countries, rates with different maturities,unemployment claims, treasury yields, etc. Because different indicators have different units,inception dates, and update frequencies, careful preprocessing is employed to align and transformthem before further analysis. It implies that the distribution depicting the same financial regimeis evolving temporally. A dynamic data-driven perspective of defining the regimes and describingtheir characteristics, as which is addressed with R2-RD, is more suitable than the common practiceof fixed thresholding. When a regime detection model is trained for the first time, the model isinitialized with sufficient data so that the fitted regime distributions contain enough information.Using the model selection process (as described previously), the candidate model families includeHMM with Gaussian observation model and HMM with Gaussian mixture observation model.There is one hyper-parameter to decide within each family of models: the number of regimes,which ranges from 2 to 10, for example. The BIC score can reveal that for all choices of hyper-parameters, Gaussian HMM has a smaller BIC score; hence a better model to describe the data.Then, four regimes have the highest Silhouette score within the Gaussian HMM family. Thus, themodel selection procedure suggests that Gaussian HMM with 4 regimes is the best model for theinitial training of macroeconomic regimes. Because regime detection models with unsupervisedmethods like HMM do not provide labeling for each regime, post-fitting analysis is utilized to gaininsights and investigate the characteristics of the regimes.

[0111] As for regime evolution, after fitting the regime detection model “rollingly,” atrajectory of regime characteristics can be acquired. This trajectory describes the regimedistributions in each retraining, and hence, provides rich information regarding the evolution ofsome defining characteristics of the regimes. Thus, the trajectory can be summarized, and thedynamics behind regime evolution can be examined. {τ1, . . . , τm} ⊂ {1, . . . , T} denotes the setof time indices of retraining. The τ-th retraining of Gaussian HMM model, which classifies allobservations before τ as a sequence of regimes {r1, . . . , rτ} where each regime rt is an element ofthe set of regimes J, is considered. With this sequence of regimes, the performance of any measuresunder a specific regime can be summarized. denotes the time series of a measure ofAttorney Docket No.: P10375PC00 Patentinterest, say the Real GDP (QoQ). Then the average Real GDP (QoQ) under regime j found by theτ-th Gaussian HMM is determined as follows:

[0112] The average Real GDP (QoQ) is examined under regime j evolves across time bylooking at the sequence , which can be visualized for four prevalentmacroeconomic indicators: Real,CPI (YoY), Unemployment Rate, and 10 YearTreasury Yield.

[0113] The R2-RD scheme can also be applied to the futures market datasets – e.g., thefollowing major types of futures contracts with minute-level price series: Gold, U.S. Treasury(UST), e-mini (S&P 500), Crude oil, and EURUSD. Under this scenario, to identify the regimesrather than simply capturing the long-term growth trend, a 1-day rolling annual returns iscalculated for each futures; the output is passed to R2-RD. It is noted that various rolling windows,ranging from minute level to hour level, does not significantly affect the results. Because futuresdata are much more volatile than macroeconomic indicators, the impact of smoothing on regimeidentification is examined. Accordingly, the heat equation smoothing, which will not create lagsin the data, is utilized.

[0114] The R2-RD scheme can also be applied to the futures market datasets – e.g., thefollowing major types of futures contracts with minute-level price series: Gold, U.S. Treasury(UST), e-mini (S&P 500), Crude oil, and EURUSD. Under this scenario, to identify the regimesrather than simply capturing the long-term growth trend, a 1-day rolling annual returns iscalculated for each futures; the output is passed to R2-RD. It is noted that various rolling windows,ranging from minute level to hour level, does not significantly affect the results. Because futuresdata are much more volatile than macroeconomic indicators, the impact of smoothing on regimeidentification is examined. Accordingly, the heat equation smoothing, which will not create lagsin the data, is utilized.

[0115] As explained above, the R2-RD framework can be readily applied to dynamicallyidentifying financial regimes. The statistical and geometric-based model selection frameworkAttorney Docket No.: P10375PC00 Patentautomatically selects the model family that best describes the data and the model within the familythat best separates the regimes. This mitigates the issue of falsely making unrealistic assumptionsabout the data generation process. By utilizing temporal ensemble, label assignment, and athreshold policy, R2-RD can identify a stable trajectory of evolving regime dynamics and theemergence of new regimes. Utilizing R2-RD, strong empirical evidence of nonstationary marketdynamics in the macroeconomic environment and the futures market can be provided; this isdifficult to identify before without the recursive training framework of R2-RD. The identifiedmacroeconomic regimes also demonstrate separation among mutual fund performance, suggestingstrong potential for regime-based asset management, including security selection and portfolioconstruction strategies.

[0116] In certain embodiments, Robust Rolling Regime Detection (R2-RD) is a HiddenMarkov Model utilizing the centroids of R2K-Means (W = 1) as a stable starting location in whichthe Gaussian distributions are fitted. The regime results of Naive K-Means are highly unstable andfrequently change due to random initialization and cluster mislabeling, rendering the regimeprobabilities unsuitable for financial application. The regime classifications of R2-RD and R2K-Means are much more chronologically stable and well separated. R2K-Means produces a timeseries of centroids primarily influenced by space with a small amount of path dependency. Sincecentroids are themselves a time series, distance metrics more applicable to time series can beutilized including Dynamic Time Warping. It is contemplated that alterations to R2K-Means canbe made to make the centroids more path driven, allowing for a longer rolling windows to producepath-related clusters across time. Additional processing steps could be added to ensure the rollingwindow of centroids do not overlap to mitigate unconnected decision boundaries. R2K-Meanscould be adapted to a variable number of k clusters throughout time to better fit the data at each tincrement.

[0117] FIGs. 5A-5H are diagrams of a graphical user interface (GUI) relating to datamanagement processes performed by the AI-based decision support platform of FIG. 1, accordingto one embodiment. GUIs 501, 503, and 505 permit the user to specify the different types ofsources of data to be collected. Such sources can be identified and selected based on asset types,macroeconomic factors, media, or even proprietary data. Via GUIs 507 and 509, the platform 101can perform predictions on the selected data sources using various models: historic, predicted,Attorney Docket No.: P10375PC00 Patentstress testing, and / or backtesting for example. According to one embodiment, GUI 511 supportsdrilling down to permit user input of various aspects of mutual funds. GUI 513 illustrates theclustering feature of platform 101. Additionally, with GUI 515, different views of the dataprocessing of clusters are supported.

[0118] FIGs. 6A-6D are diagrams of a GUI relating to asset allocation processes performed bythe AI-based decision support platform of FIG. 1, according to one embodiment. Namely, throughGUIs 601-605, different filters can be applied regarding how asset allocation can be executed.GUI 607 presents performance data associated with a selected asset or cluster of assets.

[0119] FIGs. 7A-7C are diagrams of a GUI relating to security selection processes performedby the AI-based decision support platform of FIG. 1, according to one embodiment. GUIs 701and 703, by way of example, provides specification of various data filters for an asset. Results ofthe selection process is presented in GUI 705.

[0120] FIG. 8 illustrates an example neural network 801 (e.g., an example of the AI engine211 implementing a machine learning model) that has an architecture including an input layer 803comprising one or more input neurons 805, one or more hidden neuronal layers 807 comprisingone or more hidden neurons 809, and an output layer 811 comprising one or more output neurons.In one embodiment, the architecture of the neural network 801 refers to the number of inputneurons 805, the number of neuronal layers 807, the number of hidden neurons 809 in the neuronallayers 807, the number of output neurons, or a combination thereof. In addition, the architecturecan refer to the activation function used by the neurons, the loss functions applied to train theneural network 801, parameters indicating whether the layers are fully connected (e.g., all neuronsof one layer are connected to all neurons of another layer) or partially connected, and / or otherequivalent characteristics, parameters, or properties of the neurons 805 / 809 / 813, neuronal layers807, or neural network 801. Although the various embodiments described herein are discussedwith respect to a neural network 801, it is contemplated that the various embodiments describedherein are applicable to any type of machine learning model 109 that can be migrated betweendifferent architectures.

[0121] In one embodiment, the progressive path migrates an old architecture of a machinelearning model 109 into a new architecture by incrementally adding and removing single neuronsAttorney Docket No.: P10375PC00 Patentor neuronal layers, or smoothly changing activation functions in a fashion which does not affectperformance of the machine learning model 109 by more than a designated performance changethreshold. For example, a user may wish to migrate a machine learning model 109 from anarchitecture that has three hidden neuronal layers 807 with four hidden neurons 809 in each layerto a new architecture that has four hidden neuronal layers 807 with four hidden neurons 809 each.The machine learning model 109 has been trained using the old architecture for a significant periodof time. To advantageously preserve the training already performed and maintain modelperformance at a target level, the system 100 can construct a progressive path with four steps thatincremental adds one hidden neuron 809 to the new neuronal layer 807 at each step until the fullnew neuronal layer 807 is added. In other words, while the machine learning model 109 of themachine learning system 107 is being trained, a new technical solution or architecture may bediscovered that can provide improvements to the machine learning model 109 or system 107. Theninstead of replacing the old system architecture in a cut-off fashion, the system 100 can constructincremental steps that can be used to progressively migrate the existing trained machine learningmodel 109 to avoid catastrophic degradation of the trained machine learning model 109’sperformance.

[0122] In one embodiment, while the progressive migration is being done, the training processcontinues. In this way, the newly added neurons learn relatively quickly their new roles in themachine learning model 109 as their context environment consists of neuronal layers 807 whichalready know their jobs (e.g., neuronal layers 807 with neurons 809 that have undergone at leastsome training). After migration the resulting machine learning model 109 has incorporated expertknowledge from the old architecture, but has a new architecture, new technologies incorporated,and / or the like which can potentially improve the performance and learning of the machinelearning model 109 in the future. Accordingly, the embodiments of the system 100 describedherein provide technical advantages including, but not limited to, providing long-lived machinelearning systems 107 that can be trained better while incorporating new advances in machinelearning technologies (e.g., neural network technologies).

[0123] The processes described herein for providing decision support may be advantageouslyimplemented via software, hardware, firmware or a combination of software and / or firmwareand / or hardware. For example, the processes described herein, may be advantageouslyAttorney Docket No.: P10375PC00 Patentimplemented via processor(s), Digital Signal Processing (DSP) chip, an Application SpecificIntegrated Circuit (ASIC), Field Programmable Gate Arrays (FPGAs), etc. Such exemplaryhardware for performing the described functions is detailed below.

[0124] FIG. 9 illustrates a computer system 900 upon which various embodiments of theinvention may be implemented. Although computer system 900 is depicted with respect to aparticular device or equipment, it is contemplated that other devices or equipment (e.g., networkelements, servers, etc.) within FIG. 9 can deploy the illustrated hardware and components ofsystem 900. Computer system 900 is programmed (e.g., via computer program code orinstructions) to provide decision support as described herein and includes a communicationmechanism such as a bus 910 for passing information between other internal and externalcomponents of the computer system 900. Information (also called data) is represented as a physicalexpression of a measurable phenomenon, typically electric voltages, but including, in otherembodiments, such phenomena as magnetic, electromagnetic, pressure, chemical, biological,molecular, atomic, sub-atomic and quantum interactions. For example, north and south magneticfields, or a zero and non-zero electric voltage, represent two states (0, 1) of a binary digit (bit).Other phenomena can represent digits of a higher base. A superposition of multiple simultaneousquantum states before measurement represents a quantum bit (qubit). A sequence of one or moredigits constitutes digital data that is used to represent a number or code for a character. In someembodiments, information called analog data is represented by a near continuum of measurablevalues within a particular range. Computer system 900, or a portion thereof, constitutes a meansfor performing one or more steps of the processes described herein, including that of FIG. 3.

[0125] A bus 910 includes one or more parallel conductors of information so that informationis transferred quickly among devices coupled to the bus 910. One or more processors 902 forprocessing information are coupled with the bus 910.

[0126] A processor (or multiple processors) 902 performs a set of operations on informationas specified by computer program code related to providing decision support. The computerprogram code is a set of instructions or statements providing instructions for the operation of theprocessor and / or the computer system to perform specified functions. The code, for example, maybe written in a computer programming language that is compiled into a native instruction set ofAttorney Docket No.: P10375PC00 Patentthe processor. The code may also be written directly using the native instruction set (e.g., machinelanguage). The set of operations include bringing information in from the bus 910 and placinginformation on the bus 910. The set of operations also typically include comparing two or moreunits of information, shifting positions of units of information, and combining two or more unitsof information, such as by addition or multiplication or logical operations like OR, exclusive OR(XOR), and AND. Each operation of the set of operations that can be performed by the processoris represented to the processor by information called instructions, such as an operation code of oneor more digits. A sequence of operations to be executed by the processor 902, such as a sequenceof operation codes, constitute processor instructions, also called computer system instructions or,simply, computer instructions. Processors may be implemented as mechanical, electrical,magnetic, optical, chemical, or quantum components, among others, alone or in combination.

[0127] Computer system 900 also includes a memory 904 coupled to bus 910. The memory904, such as a random access memory (RAM) or any other dynamic storage device, storesinformation including processor instructions for providing real-time data analysis to supportdecision making. Dynamic memory allows information stored therein to be changed by thecomputer system 900. RAM allows a unit of information stored at a location called a memoryaddress to be stored and retrieved independently of information at neighboring addresses. Thememory 904 is also used by the processor 902 to store temporary values during execution ofprocessor instructions. The computer system 900 also includes a read only memory (ROM) 906or any other static storage device coupled to the bus 910 for storing static information, includinginstructions, that is not changed by the computer system 900. Some memory is composed ofvolatile storage that loses the information stored thereon when power is lost. Also coupled to bus910 is a non-volatile (persistent) storage device 908, such as a magnetic disk, optical disk or flashcard, for storing information, including instructions, that persists even when the computer system900 is turned off or otherwise loses power.

[0128] Information, including instructions for providing real-time data analysis to supportdecision making, at least in part, on analysis of collected information, is provided to the bus 910for use by the processor from an external input device 912, such as a keyboard containingalphanumeric keys operated by a human user, a microphone, an Infrared (IR) remote control, ajoystick, a game pad, a stylus pen, a touch screen, or a sensor. A sensor detects conditions in itsAttorney Docket No.: P10375PC00 Patentvicinity and transforms those detections into physical expression compatible with the measurablephenomenon used to represent information in computer system 900. Other external devicescoupled to bus 910, used primarily for interacting with humans, include a display device 914, suchas a vacuum fluorescent display (VFD), a liquid crystal display (LCD), a light-emitting diode(LED), an organic light-emitting diode (OLED), a quantum dot display, a virtual reality (VR)headset, a plasma screen, a cathode ray tube (CRT), or a printer for presenting text or images, anda pointing device 916, such as a mouse, a trackball, cursor direction keys, or a motion sensor, forcontrolling a position of a small cursor image presented on the display 914 and issuing commandsassociated with graphical elements presented on the display 914, and one or more camera sensors994 for capturing, recording and causing to store one or more still and / or moving images (e.g.,videos, movies, etc.) which also may comprise audio recordings. In some embodiments, forexample, in embodiments in which the computer system 900 performs all functions automaticallywithout human input, one or more of external input device 912, a display device 914 and pointingdevice 916 may be omitted.

[0129] In the illustrated embodiment, special purpose hardware, such as an application specificintegrated circuit (ASIC) 920, is coupled to bus 910. The special purpose hardware is configuredto perform operations not performed by processor 902 quickly enough for special purposes.Examples of ASICs include graphics accelerator cards for generating images for display 914,cryptographic boards for encrypting and decrypting messages sent over a network, speechrecognition, and interfaces to special external devices, such as robotic arms and medical scanningequipment that repeatedly perform some complex sequence of operations that are more efficientlyimplemented in hardware.

[0130] Computer system 900 also includes one or more instances of a communicationsinterface 970 coupled to bus 910. Communication interface 970 provides a one-way or two-waycommunication coupling to a variety of external devices that operate with their own processors,such as printers, scanners, and external disks. In general, the coupling is with a network link 978that is connected to a local network 980 to which a variety of external devices with their ownprocessors are connected. For example, communication interface 970 may be a parallel port or aserial port or a universal serial bus (USB) port on a personal computer. In some embodiments,communications interface 970 provides an information communication connection to aAttorney Docket No.: P10375PC00 Patentcorresponding type of telephone line. In some embodiments, a communication interface 970 is acable modem that converts signals on bus 910 into signals for a communication connection over acoaxial cable or into optical signals for a communication connection over a fiber optic cable. Asanother example, communications interface 970 may be a local area network (LAN) card toprovide a data communication connection to a compatible LAN, such as Ethernet. Wireless linksmay also be implemented. For wireless links, the communications interface 970 sends or receivesor both sends and receives electrical, acoustic or electromagnetic signals, including infrared andoptical signals, that carry information streams, such as digital data. For example, in wirelesshandheld devices, such as mobile telephones like cell phones, the communications interface 970includes a radio band electromagnetic transmitter and receiver called a radio transceiver. In certainembodiments, the communications interface 970 enables connection to the communicationnetwork 97 in support of the AI-based decision support platform 101.

[0131] The term “computer-readable medium” as used herein refers to any medium thatparticipates in providing information to processor 902, including instructions for execution. Sucha medium may take many forms, including, but not limited to a computer-readable storage medium(e.g., non-volatile media, volatile media), and transmission media. Non-transitory media, such asnon-volatile media, include, for example, optical or magnetic disks, such as storage device 908.Volatile media include, for example, dynamic memory 904. Transmission media include, forexample, twisted pair cables, coaxial cables, copper wire, fiber optic cables, and carrier waves thattravel through space without wires or cables, such as acoustic waves and electromagnetic waves,including radio, optical and infrared waves. Signals include man-made transient variations inamplitude, frequency, phase, polarization or other physical properties transmitted through thetransmission media. Common forms of computer-readable media include, for example, a floppydisk, a flexible disk, hard disk, magnetic tape, any other magnetic medium, a CD-ROM, CDRW,DVD, any other optical medium, punch cards, paper tape, optical mark sheets, any other physicalmedium with patterns of holes or other optically recognizable indicia, a RAM, a PROM, anEPROM, a FLASH-EPROM, an EEPROM, a flash memory, any other memory chip or cartridge,a carrier wave, or any other medium from which a computer can read. The term computer-readablestorage medium is used herein to refer to any computer-readable medium except transmissionmedia.Attorney Docket No.: P10375PC00 Patent

[0132] Logic encoded in one or more tangible media includes one or both of processorinstructions on a computer-readable storage media and special purpose hardware, such as ASIC920.

[0133] Network link 978 typically provides information communication using transmissionmedia through one or more networks to other devices that use or process the information. Forexample, network link 978 may provide a connection through local network 980 to a host computer982 or to equipment 984 operated by an Internet Service Provider (ISP). ISP equipment 984 inturn provides data communication services through the public, world-wide packet-switchingcommunication network of networks now commonly referred to as the Internet 990.

[0134] A computer called a server host 992 connected to the Internet hosts a process thatprovides a service in response to information received over the Internet. For example, server host992 hosts a process that provides information representing video data for presentation at display914. It is contemplated that the components of system 900 can be deployed in variousconfigurations within other computer systems, e.g., host 982 and server 992.

[0135] At least some embodiments of the invention are related to the use of computer system900 for implementing some or all of the techniques described herein. According to oneembodiment of the invention, those techniques are performed by computer system 900 in responseto processor 902 executing one or more sequences of one or more processor instructions containedin memory 904. Such instructions, also called computer instructions, software and program code,may be read into memory 904 from another computer-readable medium such as storage device908 or network link 978. Execution of the sequences of instructions contained in memory 904causes processor 902 to perform one or more of the method steps described herein. In alternativeembodiments, hardware, such as ASIC 920, may be used in place of or in combination withsoftware to implement the invention. Thus, embodiments of the invention are not limited to anyspecific combination of hardware and software, unless otherwise explicitly stated herein.

[0136] The signals transmitted over network link 978 and other networks throughcommunications interface 970, carry information to and from computer system 900. Computersystem 900 can send and receive information, including program code, through the networks 980,990 among others, through network link 978 and communications interface 970. In an exampleAttorney Docket No.: P10375PC00 Patentusing the Internet 990, a server host 992 transmits program code for a particular application,requested by a message sent from computer 900, through Internet 990, ISP equipment 984, localnetwork 980 and communications interface 970. The received code may be executed by processor902 as it is received, or may be stored in memory 904 or in storage device 908 or any other non-volatile storage for later execution, or both. In this manner, computer system 900 may obtainapplication program code in the form of signals on a carrier wave.

[0137] Various forms of computer readable media may be involved in carrying one or moresequence of instructions or data or both to processor 902 for execution. For example, instructionsand data may initially be carried on a magnetic disk of a remote computer such as host 982. Theremote computer loads the instructions and data into its dynamic memory and sends theinstructions and data over a telephone line using a modem. A modem local to the computer system900 receives the instructions and data on a telephone line and uses an infra-red transmitter toconvert the instructions and data to a signal on an infra-red carrier wave serving as the networklink 978. An infrared detector serving as communications interface 970 receives the instructionsand data carried in the infrared signal and places information representing the instructions and dataonto bus 910. Bus 910 carries the information to memory 904 from which processor 902 retrievesand executes the instructions using some of the data sent with the instructions. The instructionsand data received in memory 904 may optionally be stored on storage device 908, either before orafter execution by the processor 902.

[0138] FIG. 10 illustrates a chip set or chip 1000 upon which various embodiments of theinvention may be implemented. Chip set 1000 is programmed to the processes (e.g., FIG. 3) asdescribed herein and includes, for instance, the processor and memory components described withrespect to FIG. 9 incorporated in one or more physical packages (e.g., chips). By way of example,a physical package includes an arrangement of one or more materials, components, and / or wireson a structural assembly (e.g., a baseboard) to provide one or more characteristics such as physicalstrength, conservation of size, and / or limitation of electrical interaction. It is contemplated that incertain embodiments the chip set 1000 can be implemented in a single chip. It is furthercontemplated that in certain embodiments the chip set or chip 1000 can be implemented as a single“system on a chip.” It is further contemplated that in certain embodiments a separate ASIC wouldnot be used, for example, and that all relevant functions as disclosed herein would be performedAttorney Docket No.: P10375PC00 Patentby a processor or processors. Chip set or chip 1000, or a portion thereof, constitutes a means forperforming one or more steps of providing user interface navigation information associated withthe availability of functions. Chip set or chip 1000, or a portion thereof, constitutes a means forperforming one or more steps of providing decision support.

[0139] In one embodiment, the chip set or chip 1000 includes a communication mechanismsuch as a bus 1001 for passing information among the components of the chip set 1000. Aprocessor 1003 has connectivity to the bus 1001 to execute instructions and process informationstored in, for example, a memory 1005. The processor 1003 may include one or more processingcores with each core configured to perform independently. A multi-core processor enablesmultiprocessing within a single physical package. Examples of a multi-core processor includetwo, four, eight, or greater numbers of processing cores. Alternatively or in addition, the processor1003 may include one or more microprocessors configured in tandem via the bus 1001 to enableindependent execution of instructions, pipelining, and multithreading. The processor 1003 mayalso be accompanied with one or more specialized components to perform certain processingfunctions and tasks such as one or more digital signal processors (DSP) 1007, or one or moreapplication-specific integrated circuits (ASIC) 1009. A DSP 1007 typically is configured toprocess real-world signals (e.g., sound) in real time independently of the processor 1003.Similarly, an ASIC 1009 can be configured to performed specialized functions not easilyperformed by a more general purpose processor. Other specialized components to aid inperforming the inventive functions described herein may include one or more field programmablegate arrays (FPGA), one or more controllers, or one or more other special-purpose computer chips.

[0140] In one embodiment, the chip set or chip 1000 includes merely one or more processorsand some software and / or firmware supporting and / or relating to and / or for the one or moreprocessors.

[0141] The processor 1003 and accompanying components have connectivity to the memory1005 via the bus 1001. The memory 1005 includes both dynamic memory (e.g., RAM, magneticdisk, writable optical disk, etc.) and static memory (e.g., ROM, CD-ROM, etc.) for storingexecutable instructions that when executed perform the inventive steps described herein to provideAttorney Docket No.: P10375PC00 Patentproviding decision support. The memory 1005 also stores the data associated with or generatedby the execution of the inventive steps.

[0142] FIG. 11 is a diagram of exemplary components of a mobile terminal (e.g., handset) forcommunications, which is capable of operating in the system of FIG. 1, according to oneembodiment. In some embodiments, mobile terminal 1101, or a portion thereof, constitutes ameans for performing one or more steps of the described processes. Generally, a radio receiver isoften defined in terms of front-end and back-end characteristics. The front-end of the receiverencompasses all of the Radio Frequency (RF) circuitry whereas the back-end encompasses all ofthe base-band processing circuitry. As used in this application, the term “circuitry” refers to both:(1) hardware-only implementations (such as implementations in only analog and / or digitalcircuitry), and (2) to combinations of circuitry and software (and / or firmware) (such as, ifapplicable to the particular context, to a combination of processor(s), including digital signalprocessor(s), software, and memory(ies) that work together to cause an apparatus, such as a mobilephone or server, to perform various functions). This definition of “circuitry” applies to all uses ofthis term in this application, including in any claims. As a further example, as used in thisapplication and if applicable to the particular context, the term “circuitry” would also cover animplementation of merely a processor (or multiple processors) and its (or their) accompanyingsoftware / or firmware. The term “circuitry” would also cover if applicable to the particular context,for example, a baseband integrated circuit or applications processor integrated circuit in a mobilephone or a similar integrated circuit in a cellular network device or other network devices.

[0143] Pertinent internal components of the telephone include a Main Control Unit (MCU)1103, a Digital Signal Processor (DSP) 1105, and a receiver / transmitter unit including amicrophone gain control unit and a speaker gain control unit. A main display unit 1107 providesa display to the user in support of various applications and mobile terminal functions that performor support the steps of providing decision support. The display 1107 includes display circuitryconfigured to display at least a portion of a user interface of the mobile terminal (e.g., mobiletelephone). Additionally, the display 1107 and display circuitry are configured to facilitate usercontrol of at least some functions of the mobile terminal. An audio function circuitry 1109 includesa microphone 1111 and microphone amplifier that amplifies the speech signal output from theAttorney Docket No.: P10375PC00 Patentmicrophone 1111. The amplified speech signal output from the microphone 1111 is fed to acoder / decoder (CODEC) 1113.

[0144] A radio section 1115 amplifies the power and converts frequency in order tocommunicate with a base station, which is included in a mobile communication system, viaantenna 1117. The power amplifier (PA) 1119 and the transmitter / modulation circuitry areoperationally responsive to the MCU 1103, with an output from the PA 1119 coupled to theduplexer 1121 or circulator or antenna switch, as known in the art. The PA 1119 also couples toa battery interface and power control unit 1120.

[0145] In use, a user of mobile terminal 1101 speaks into the microphone 1111 and his or hervoice along with any detected background noise is converted into an analog voltage. The analogvoltage is then converted into a digital signal through the Analog to Digital Converter (ADC) 1123.The control unit 1103 routes the digital signal into the DSP 1105 for processing therein, such asspeech encoding, channel encoding, encrypting, and interleaving. In one embodiment, theprocessed voice signals are encoded, by units not separately shown, using a cellular transmissionprotocol such as enhanced data rates for global evolution (EDGE), general packet radio service(GPRS), global system for mobile communications (GSM), Internet protocol multimediasubsystem (IMS), universal mobile telecommunications system (UMTS), etc., as well as any othersuitable wireless medium, e.g., microwave access (WiMAX), Long Term Evolution (LTE)networks, code division multiple access (CDMA), wideband code division multiple access(WCDMA), wireless fidelity (WiFi), satellite, and the like, or any combination thereof.

[0146] The encoded signals are then routed to an equalizer 1125 for compensation of anyfrequency-dependent impairments that occur during transmission though the air such as phase andamplitude distortion. After equalizing the bit stream, the modulator 1127 combines the signal withan RF signal generated in the RF interface 1129. The modulator 1127 generates a sine wave byway of frequency or phase modulation. In order to prepare the signal for transmission, an up-converter 1131 combines the sine wave output from the modulator 1127 with another sine wavegenerated by a synthesizer 1133 to achieve the desired frequency of transmission. The signal isthen sent through a PA 1119 to increase the signal to an appropriate power level. In practicalsystems, the PA 1119 acts as a variable gain amplifier whose gain is controlled by the DSP 1105Attorney Docket No.: P10375PC00 Patentfrom information received from a network base station. The signal is then filtered within theduplexer 1121 and optionally sent to an antenna coupler 1135 to match impedances to providemaximum power transfer. Finally, the signal is transmitted via antenna 1117 to a local base station.An automatic gain control (AGC) can be supplied to control the gain of the final stages of thereceiver. The signals may be forwarded from there to a remote telephone which may be anothercellular telephone, any other mobile phone or a land-line connected to a Public SwitchedTelephone Network (PSTN), or other telephony networks.

[0147] Voice signals transmitted to the mobile terminal 1101 are received via antenna 1117and immediately amplified by a low noise amplifier (LNA) 1137. A down-converter 1139 lowersthe carrier frequency while the demodulator 1141 strips away the RF leaving only a digital bitstream. The signal then goes through the equalizer 1125 and is processed by the DSP 1105. ADigital to Analog Converter (DAC) 1143 converts the signal and the resulting output is transmittedto the user through the speaker 1145, all under control of a Main Control Unit (MCU) 1103 whichcan be implemented as a Central Processing Unit (CPU).

[0148] The MCU 1103 receives various signals including input signals from the keyboard1147. The keyboard 1147 and / or the MCU 1103 in combination with other user input components(e.g., the microphone 1111) comprise a user interface circuitry for managing user input. The MCU1103 runs a user interface software to facilitate user control of at least some functions of the mobileterminal 1101 to provide decision support. The MCU 1103 also delivers a display command anda switch command to the display 1107 and to the speech output switching controller, respectively.Further, the MCU 1103 exchanges information with the DSP 1105 and can access an optionallyincorporated SIM card 1149 and a memory 1151. In addition, the MCU 1103 executes variouscontrol functions required of the terminal. The DSP 1105 may, depending upon theimplementation, perform any of a variety of conventional digital processing functions on the voicesignals. Additionally, DSP 1105 determines the background noise level of the local environmentfrom the signals detected by microphone 1111 and sets the gain of microphone 1111 to a levelselected to compensate for the natural tendency of the user of the mobile terminal 1101.

[0149] The CODEC 1113 includes the ADC 1123 and DAC 1143. The memory 1151 storesvarious data including call incoming tone data and is capable of storing other data including musicAttorney Docket No.: P10375PC00 Patentdata received via, e.g., the global Internet. The software module could reside in RAM memory,flash memory, registers, or any other form of writable storage medium known in the art. Thememory device 1151 may be, but not limited to, a single memory, CD, DVD, ROM, RAM,EEPROM, optical storage, magnetic disk storage, flash memory storage, or any other non-volatilestorage medium capable of storing digital data.

[0150] An optionally incorporated SIM card 1149 carries, for instance, important information,such as the cellular phone number, the carrier supplying service, subscription details, and securityinformation. The SIM card 1149 serves primarily to identify the mobile terminal 1101 on a radionetwork. The card 1149 also contains a memory for storing a personal telephone number registry,text messages, and user specific mobile terminal settings.

[0151] Further, one or more camera sensors 1153 may be incorporated onto the mobile station1101 wherein the one or more camera sensors may be placed at one or more locations on the mobilestation. Generally, the camera sensors may be utilized to capture, record, and cause to store oneor more still and / or moving images (e.g., videos, movies, etc.) which also may comprise audiorecordings.

[0152] While the invention has been described in connection with a number of embodimentsand implementations, the invention is not so limited but covers various obvious modifications andequivalent arrangements, which fall within the purview of the appended claims. Although featuresof the invention are expressed in certain combinations among the claims, it is contemplated thatthese features can be arranged in any combination and order.

Claims

Attorney Docket No.: P10375PC00 PatentCLAIMSWHAT IS CLAIMED IS:

1. A method comprising:generating a data-set having,a first dimension representing a rolling window in a time series,a second dimension representing number of financial instruments,a third dimension representing number of observations through a predetermined timeperiod, anda fourth dimension representing number of features;clustering the data-set;setting cluster size and rolling window length;applying K-Means algorithm to the data-set to establish a plurality of centroids;for each of the plurality of centroids, selecting closest data point in a current data-set,wherein each centroid has at least one data point;iteratively applying the K-Means algorithm according to the selected data points until thecorresponding centroids converge; anddetermining scores for the centroids to assess separation of the clusters, wherein the scoresare utilized to determine performance of the financial instruments.

2. The method of claim 1, further comprising:assigning a closest data point from a previous period for each of the centroids to yield newlocations for each of the centroids.

3. The method of claim 1, wherein the rolling window length is set so that the clusters arenonlinear and connected.

4. The method of claim 1, further comprising:Attorney Docket No.: P10375PC00 Patentsorting the centroids in ascending order based on number of data points in each of thecorresponding centroids.

5. The method of claim 1, further comprising:applying a threshold policy to determine number of regimes associated with the data set.

6. The method of claim 1, wherein the financial instruments include a plurality of funds.

7. A system comprising:a memory configured to store computer-executable instructions; andone or more processors configured to execute the instructions to:generate a data-set having,a first dimension representing a rolling window in a time series,a second dimension representing number of financial instruments,a third dimension representing number of observations through a predetermined timeperiod, anda fourth dimension representing number of features;cluster the data-set;set cluster size and rolling window length;apply K-Means algorithm to the data-set to establish a plurality of centroids;for each of the plurality of centroids, selecting closest data point in a current data-set,wherein each centroid has at least one data point;iteratively apply the K-Means algorithm according to the selected data points until thecorresponding centroids converge; anddetermine scores for the centroids to assess separation of the clusters, wherein the scoresare utilized to determine performance of the financial instruments.

8. The system of claim 7, wherein the one or more processors are further configured toexecute the instructions to:Attorney Docket No.: P10375PC00 Patentassign a closest data point from a previous period for each of the centroids to yield newlocations for each of the centroids.

9. The system of claim 7, wherein the rolling window length is set so that the clusters arenonlinear and connected.

10. The system of claim 7, wherein the one or more processors are further configured toexecute the instructions to:sort the centroids in ascending order based on number of data points in each of thecorresponding centroids.

11. The system of claim 7, wherein the one or more processors are further configured toexecute the instructions to:apply a threshold policy to determine number of regimes associated with the data set.

12. The system of claim 7, wherein the financial instruments include a plurality of funds.

13. A method comprising:identifying a plurality of regimes corresponding to statistical models;computing a plurality of scores for the statistical models, wherein each of the plurality ofscores measures similarity of a data point to a corresponding cluster against dissimilarityof another cluster;comparing the statistical models using marginal likelihood;selecting, for machine learning, one of the statistical models based on the plurality of scores;and selecting one of a plurality of financial instruments based on the selected statistical model.

14. The method of claim 13, further comprising:calculating posterior odds for the selection of the statistical models.Attorney Docket No.: P10375PC00 Patent15. The method of claim 14, further comprising:determining a Bayesian Information Criterion (BIC) score to approximate the marginallikelihood, wherein the BIC score is used for the comparison of the statistical models.

16. The method of claim 13, wherein the statistical models are based a Hidden MarkovModel, the method further comprising:utilizing a plurality of centroids output from a K-Means algorithm as a starting location forthe Hidden Markov Model.

17. The method of claim 16, wherein the plurality of centroids provide rolling windows toproduce path-related clusters across time.

18. The method of claim 17, wherein the rolling windows for the plurality of centroids arenon-overlapping.

19. The method of claim 13, wherein the regimes relate to either a macroeconomic marketdata or futures market data.

20. The method of claim 19, wherein the financial instruments include securities.

Images