A system for topologically reconfiguring classical technical analysis indicators into modular neural networks

Abstract

The present invention relates to a system for topological reconstruction of classical technical analysis indicators into modular neural networks, for topological reconstruction of classical technical trading indicators, including moving average (MA), moving average convergence / divergence (MACD), relative strength index (RSI), commodity channel index (CCI), rate of change (ROC), and random oscillator indicator, as well as any other indicator that can be represented structurally through the defined operators of the present TIN system (see Fig. 1). The TIN system addresses the lack of interpretability and adaptability of traditional indicator models by converting their mathematical structure into a reconfigurable network architecture. Each sub-network (indicator network) corresponds to a technical indicator, whose logical operation structure is based on the combination of these operators, including but not limited to weighted average, difference, normalization, division, clipping mechanism, or minimum / maximum value determination. The input layer supports structured and unstructured, especially financial market-related data. The weights can be initialized analytically or adjusted dynamically using various learning methods, including but not limited to reinforcement learning. The present system is oriented towards algorithmic securities trading, focusing on transparency, interpretability, and real-time adaptability.

Description

Technical Field

[0001] This invention relates to a method and system for processing financial indicators using neural network technology, and particularly to an adaptive, interpretable neural architecture model for mapping classical technical analysis indicators, including but not limited to moving averages (MA), moving average converging / converging divergence (MACD), relative strength index (RSI), commodity channel index (CCI), rate of change (ROC), and stochastic oscillators, as long as their mathematical structure can be transformed into a layer-based neural network topology through transferable operators—as defined by the Technical Indicator Network (TIN system) introduced in this invention.

[0002] This technical field is an interdisciplinary field, specifically covering the following categories:

[0003] Algorithmic trading (algorithmic securities trading);

[0004] Deep learning and neural networks;

[0005] Signal processing in financial markets;

[0006] Explainable artificial intelligence in the financial sector.

[0007] Specifically, the present invention relates to the conception, construction, and use of so-called TIN systems, which are constructed as formal, interpretable neural network architectures for representing and extending rule-based technical analysis indicators. These so-called TIN systems consist of standardized layer modules whose architecture and weights are based on the mathematical structure of the original indicators.

[0008] The starting point of this invention is the observation that many widely used technical trading indicators have a mathematical structure equivalent to a simple neural network layer. Based on this, a framework is constructed in which each classic indicator is mapped by a neural subnetwork (Indicator Network, IN), which is characterized by its topology, weight matrix, and, where necessary, reinforcement learning methods.

[0009] Therefore, this invention belongs to the fields of machine learning, deep learning modeling, and signal-based decision architectures in finance. Specifically, it enables classic trading indicators to be practically applied within modern AI-supported systems through structure-fidelity, interpretable, and initializeable neural networks.

[0010] Furthermore, this invention relates to applications in quantitative analysis, automated portfolio management, signal-driven trade execution, and the development of reusable AI components for financial applications (financial informatics). TIN represents a structured approach for transforming technical indicators into scalable, reconfigurable networks that can be extended through domain-specific functions. This allows for the dynamic adjustment and training of mature indicators using AI methods and their integration into automated trading systems. Therefore, this invention possesses methodological and technical depth in network design and is highly relevant to practical applications in institutional trading systems where transparency, stability, and interpretability are crucial. Background Technology

[0011] Existing technologies disclose various applications of deep learning models in financial market environments, particularly Long Short-Term Memory (LSTM), Convolutional Neural Networks (CNN), Transformers, and reinforcement learning for predicting price movements and optimizing portfolios. However, in these methods, technical indicators are mostly treated as simple input vectors without considering their internal computational logic or structurally interpretable representations. Therefore, current practices are limited to using technical indicators in a black-box manner, failing to leverage their rule-based origins.

[0012] However, the following problems exist in practice: Although many classic technical trading indicators originated in the last century, they are still actively used today, providing robust and interpretable signals for different market conditions. Their rule-based construction and empirically validated effectiveness constitute a valuable knowledge base, which has not yet been systematically utilized or continued in modern artificial intelligence systems. The challenge lies in how to transform this existing knowledge into up-to-date, machine-interpretable models without losing the mature mathematical structure. Specifically:

[0013] 1. Classic trading indicators lack structured mappings within neural networks;

[0014] 2. The interpretability of the decision-making basis of neural trading models is low;

[0015] 3. The optimization cost of parameterizing technical indicators is relatively high;

[0016] 4. The existing architecture design has low reusability or modularity;

[0017] 5. There are limitations to integrating additional data sources such as sentiment analysis, trading volume, or cross-asset relationships into classic indicators;

[0018] 6. Insufficient reproducibility due to the non-determinism of the training process;

[0019] 7. Modern deep learning models have high energy consumption and computational overhead. Summary of the Invention

[0020] The technical problem to be solved

[0021] The purpose of this invention is to provide a technical method and a system employing a TIN (Trading Indicator Network) architecture, which can structurally and semantically transfer existing, effective trading indicators into a reconstructed neural network. In this process, the original logic of the indicator is mapped through a topological network structure, and the initial weights are set according to the original parameters. Thus, the core functionality of the classical method is preserved and transformed into a scalable and learnable network architecture. This architecture serves as the methodological and structural foundation for systematically transforming existing rule-based trading indicators into trainable and interpretable neural networks. This system is not a purely mathematical redefinition, but a technical architecture built upon mature indicator logic, utilizing its structural and operational characteristics through modern artificial intelligence. This invention thus solves the technical problems of insufficient transparency, adaptability, and modularity when using technical indicators in artificial intelligence and algorithmic trading environments.

[0022] Technical solution

[0023] This invention relates to a system for processing technical financial indicators, based on a modular neural network structure. The TIN system transforms classic technical trading indicators—such as MA, MACD, RSI, CCI, ROC, or stochastic oscillators—into reconfigurable network layers. Each sub-network (Indicator Network, IN) maps a specific indicator within a topology-defined network architecture. This architecture maintains backward compatibility with the original mathematical definitions and allows the model to be used directly without initial training.

[0024] An index network (IN) includes at least the following components:

[0025] Input Layer: This layer receives sequential data streams such as price time series, trading volume, order book data, sentiment vectors (e.g., generated by NLP models, rule-based classifiers, or other systems and processes), fundamental data, macroeconomic variables, or other data-driven sources, and the examples described are not exhaustive. This layer supports structured and unstructured data sources in vectorized or preprocessed form and allows for the integration of multiple data modalities from a variety of different sources.

[0026] Processing Layer: The inner structure follows mathematical logic defined by the corresponding technical specifications. It includes... Figure 1 The standardized operators include weighted averaging, differencing, division with bias regularization, amplitude limiting, minimum / maximum determination, and statistical bias calculation. The network topology is directly derived from the original formula of the index.

[0027] Intermediate layers: Used for normalization, context integration, noise reduction, pooling, or other auxiliary transformations. Any layer commonly used in deep learning and capable of being integrated into a TIN system can be used, such as bias-regularized division, layer normalization, context filters, rescaling layers, attention mechanisms, or custom normalization methods. The choice of intermediate layers is not exhaustive and can be flexibly adapted to corresponding metric logic.

[0028] Output layer: Generates scalable numerical decision quantities. These can be binary or categorical classifications (e.g., buy / hold / sell) or continuous regression outputs (e.g., rating scale, risk value). Decisions are based on threshold logic, softmax classification, regression functions, or other suitable decision-making methods, depending on the application scenario and target metric.

[0029] The initialization of weights can be performed optionally: (a) analytically, by preserving the original indicator structure and directly using the parameters specified in the original definition (e.g., a 12-period moving average in an EMA) as the initial weights; or (b) adaptively, by using reinforcement learning, supervised learning, or other machine learning methods, based on historical data, market feedback, or a measure of funding objectives.

[0030] This TIN system enables: the transformation of classical indicator functions into topologically defined neural networks; interpretable decision-making processes based on reconfigurable layer structures; adaptive weight optimization using reinforcement learning, supervised learning, or other learning-based methods, particularly considering real-world market conditions such as volatility states, capital requirements, drawdown limits, or risk budgets; modular scalability for additional data modalities (e.g., text data, order book data, event data); and a standardized, reusable architecture for deriving technical indicators based on defined operators (e.g., weighted average, differencing, division, MAD, minimum / maximum determination). Figure 1 As categorized by the system.

[0031] Furthermore, the TIN system can integrate multiple INs in parallel within a modular multi-index system. The outputs of multiple INs can be embedded into higher-level decision or analysis modules. This embedding can be done in an aggregated, adaptive, or meta-network-based manner; weighted fusion is only a special case here. This system allows existing technical analysis methods to be directly transformed into AI-supported decision structures without altering the mathematical logic. Thus, mature indices remain fully usable but become adaptively scalable, interpretable, and composable through neural representations.

[0032] Beneficial effects

[0033] This invention maintains full backward compatibility with the original indicator logic because, from a network technology perspective, many classic indicators can be interpreted as special cases of neural networks with fixed preset weights. Simultaneously, by initially parameterizing the network layers of the topological representation based on the originally defined indicator values, it solves many core initialization problems of deep learning models. This improves convergence, significantly reduces training overhead, and enables it to at least reproduce—and often surpass—the performance of classic models, while maintaining algorithmic continuity with existing trading logic.

[0034] In contrast, TIN systems, which generate representations of the MACD indicator, typically contain only a few thousand trainable parameters, while standard Transformer models require millions. This results in significantly reduced computational and data requirements, making TIN systems not only technically efficient but also resource-saving and sustainable—especially in real-time applications with high-frequency demands. Compared to general-purpose architectures like LSTM, CNN, or Transformer, TINs have significantly fewer parameters (e.g., fewer than 10,000 instead of millions), resulting in substantial reductions in energy consumption and hardware requirements, making them particularly suitable for energy-efficient (“green AI”) applications in algorithmic real-time trading environments.

[0035] The reduction in training complexity and time also directly impacts the energy and cost efficiency of developing production-ready trading systems. Therefore, the TIN system achieves significant efficiency improvements compared to complex deep learning models, making it particularly suitable for real-time environments in high-frequency trading where minimal latency, computational load, and rapid response are crucial. This invention bridges the gap between classical technical analysis and modern AI-based model architectures through a systematic and technically interpretable method of mapping indicator logic to neural network structures, filling a technological gap in the practical application of deep learning in the financial field. Attached Figure Description

[0036] Figure 1 : Operator table used for processing layers;

[0037] Figure 2 Topology of MMA (Multiple Moving Average) in TIN;

[0038] Figure 3 Example of a TIN using one-dimensional input data;

[0039] Figure 4 The topological structure of the MACD indicator in the TIN system;

[0040] Figure 5 : A one-dimensional TIN used to structurally implement a technical indicator in a TIN system;

[0041] Figure 6 The TIN system employs a scalable TIN (based on MACD) with two-dimensional input data.

[0042] Figure 7 The network structure of the ROC index in the TIN system;

[0043] Figure 8 The network structure of the RSI index in the TIN system;

[0044] Figure 9 : The network structure of the stochastic oscillation index in the TIN system;

[0045] Figure 10 The network structure of the CCI index in the TIN system.

[0046] Explanation of reference numerals in the attached figures:

[0047] 101—Weighted average layer; 102—Subtraction / addition layer; 103—Mean absolute deviation and standard deviation layer (MAD / SD layer); 104—Multiplication / division layer; 105—Upper limit / lower limit layer (O=upper limit, U=lower limit); 106—Minimum / maximum limit layer;

[0048] 201—Input layer; 202—Weighted average layer (fast); 203—Weighted average layer (slow); 204—Subtraction layer; 205—Output layer (MMA signal);

[0049] 301—One-dimensional data input layer; 302—Fast and slow weighted averaging layer; 303—Subtraction layer; 304—Output layer with multiple signals;

[0050] 401—Input layer; 402, 403—Weighted average layer (short-term / long-term); 404—Subtraction layer; 405, 406—Smoothing layer; 407—Second subtraction layer; 408—Output layer;

[0051] 501—Input Layer (Time, Price, or Volume Data); 502—Slow MA Layer; 503—Fast MA Layer; 504—Subtraction Layer; 505—Aggregation / Merger Layer; 506—Terminal Difference Layer;

[0052] 601—Two-dimensional data input layer; 602—Weighted feature extraction; 603—Subtraction layer; 604—Fusion and smoothing; 605—Output layer with decision vector;

[0053] 701—Input layer; 702, 703—Aggregation layer; 704—Division layer; 705, 706—Comparison aggregation layer; 707—Subtraction layer; 708—Output layer (ROC signal);

[0054] 801—Input Layer; 802—Aggregation Layer; 803, 804—Upper / Lower Bound Layers; 805, 806—Moving Average Layers; 807—Division Layer; 808—Output Layer;

[0055] 901—Input layer; 902, 903—Min / Maximum layer; 904, 905—Subtraction layer; 906—Division layer; 907—Smoothing layer; 908—Output layer;

[0056] 1001—Input layer; 1002, 1003, 1004—Average value layer; 1005—MAD layer; 1006—Subtraction layer; 1007—Division layer; 1008—Output layer.

[0057] Note: The following figures are taken directly from the original application documents. The boxed labels "EINGABESCHICHT" correspond to the Chinese term "Input Layer," and "AUSGABESCHICHT" corresponds to the Chinese term "Output Layer." It is recommended to replace the German labels in the figures with the corresponding Chinese characters before formal submission. In each figure, the circular node Σ represents a weighted average / summation operation, "−" represents subtraction, and "÷ / ·|·" represents division. MAD, SD, MIN, MAX, O, and U correspond to the operators described in the specification. Detailed Implementation

[0058] The invention will now be described by way of example with reference to the accompanying drawings. The input layer receives any transaction-related input data—which can be raw data (e.g., unprocessed price and volume streams) or preprocessed data (e.g., normalized price or volume values, technical indicator values, or sentiment indicators), and is not limited to financial data. It can integrate any relevant time-series data source to improve the stability and efficiency of subsequent networks. Implementation is achieved through topologically structured indicator networks (INs), which transform the mathematical structure and signal processing of the underlying indicators into a modular neural architecture. The topology used is based on… Figure 1 The table of operators defines a combination of basic operators. It also provides the correspondence between the original computational logic of the indicators and the functional processing layers of the neural network.

[0059] Figure 1 Operator table used for processing layers

[0060] Figure 1 The table provides an overview of the operators in TIN used to map classic financial indicators, along with their corresponding mathematical formulas.

[0061] 101 — Weighted Average Layer: A layer used to perform weighted averaging over a time window of length N. Each processing unit (hereinafter referred to as a neuron) within this layer represents a neuron with its own weight w. k,iThe specific weighted moving average is described by the following formula:

[0062] MA k,t = Σ i=0 N−1 w k,i ·P t−i , Σ i=0 N−1 w k,i = 1

[0063] In classic indicator design, the sum of weights corresponds to a normalization of 1, but within the TIN system framework, it can be constructed as a learnable parameter structure. Example indicators: Moving Average (MA), Moving Average Convergence Divergence (MACD), Multiple Moving Average (MMA), Commodity Channel Index (CCI), Relative Strength Index (RSI).

[0064] 102 — Subtraction / Addition Layer: A layer used to calculate the difference or sum of two input signals. Each neuron performs its own operation:

[0065] D k,t = A k,t - B k,t , S k,t = A k,t + B k,t

[0066] Example indicators: MACD, MMA, CCI, Stochastic Oscillator.

[0067] 103 — Mean Absolute Deviation and Standard Deviation Layer (MAD / SD Layer): A layer used to determine the mean absolute deviation (MAD) and standard deviation (SD) over a window of length N. Each neuron calculates for its input signal:

[0068] MAD k,t = (1 / N) Σ i=0 N−1 | P k,t−i - P̄ k |

[0069] SD k,t = √( (1 / N) Σ i=0 N−1 (P) k,t−i - P̄ k ) 2 )

[0070] Among them P̄ k = (1 / N) Σ i=0 N−1 P k,t−i

[0071] Example indicators: CCI, True Volatility (TR).

[0072] 104 — Multiplication / Division Layer: A layer used to perform multiplication and division operations between input signals. Each neuron performs its own multiplication or division operation:

[0073] M k,t = A k,t ·B k,t Div k,t = A k,t / (B) k,t + ε )

[0074] Example indicators: RSI, Rate of Change (ROC), CCI, Stochastic Oscillator.

[0075] 105 — Upper / Lower Bound Layers: Layers used to restrict values ​​to defined upper bounds O and lower bounds U. Each neuron restricts its input value using the following formula:

[0076] upper limit k,t = min( X k,t , O ), lower limit k,t = max( X k,t , U )

[0077] Example metric: RSI.

[0078] 106 — Minimum / Maximum Layer: A summary layer used to determine local extrema within a window. Each neuron determines an extremum for its input:

[0079] MIN k,t = min i=0,…,N−1 P k,t−i , MAX k,t = max i=0,…,N−1 P k,t−i

[0080] Example indicators: Stochastic Oscillator, support and resistance.

[0081] Figure 2 MMA topology in TIN

[0082] Figure 2 This illustrates the TIN topology for the MMA metric within the TIN system framework:

[0083] 201 — Input layer: Collects sequential price data streams and forwards them to multiple parallel aggregation paths.

[0084] 202 — Weighted Average Layer (Fast): According to operator 101 ( Figure 1 The weighted average layer is used to calculate the fast moving average with a shorter time window.

[0085] 203 — Weighted average layer (slow): according to operator 101 ( Figure 1 The weighted average layer is used to form a slow moving average with a longer time window.

[0086] 204 — Subtraction layer: according to operator 102 ( Figure 1 The difference layer is used to subtract the result of the slow averaging from the result of the fast averaging to produce the original signal of MMA.

[0087] 205 — Output Layer (MMA Signal): Integrates the differential signals and forwards them as trading signals to the upper-level decision-making or aggregation modules.

[0088] This explanation clarifies how the TIN system incorporates the classical computational form of MMA.

[0089] MMA t = MA t 快 - MA t 慢 = Σ i=0 N_f−1 w i (f) ·P t−i − Σ j=0 N_s−1 w j (s) ·P t−j

[0090] The algorithm is transformed into a layer-based neural network structure. The mapping method is as follows: the weighted sums MA (fast) and MA (slow) are implemented through parallel layers 202 and 203 (operator 101); the difference between the two component values ​​is implemented through layer 204 (operator 102). Thus, all the components of the formula are implemented through modular and reconfigurable layers, thereby realizing an interpretable and trainable implementation of the MMA index in the TIN system.

[0091] Figure 3 TIN example using one-dimensional input data

[0092] Figure 3 An example TIN is shown, which uses one-dimensional price data (301) or other trading-related time series data as input signals. The network structure is based on... Figure 2 The MMA topology shown is used, and the following is employed: Figure 1 The specified operators are implemented in a modular fashion.

[0093] 301 — One-dimensional data input layer: Collects sequential price data points or other relevant input signals and forwards them to parallel processing branches.

[0094] 302 — Fast and Slow Weighted Average Layers: According to Operator 101 ( Figure 1 The two weighted average layers correspond to Figure 2 The defined MA structure. Weight initialization can be performed according to rules based on the original metrics, or based on data using learning methods.

[0095] 303 — Subtraction layer: according to operator 102 ( Figure 1 The difference layer combines the results of the two average layers to represent the deviation between price trends.

[0096] 304 — Output layer with multiple signals: integrates the differential signals and can generate multiple numerical output values ​​(e.g., binary decision, regression value, risk assessment) depending on the configuration.

[0097] This example structure illustrates how to... Figure 2 The exported network extension is used for one-dimensional input. Through combination... Figure 1 The defined layers and Figure 2 The structured topology of the network constitutes a scalable and trainable model for mapping technical metrics in a neural context.

[0098] Figure 4 The topological structure of the MACD indicator in the TIN system

[0099] Figure 4 This diagram illustrates the network topology representing the classic MACD indicator within a TIN. The structure is based on sequences derived from moving averages and differencing operations, and utilizes… Figure 1 The defined operator.

[0100] 401 — Input layer: Collects sequential price data or similar transaction information.

[0101] 402 and 403 — Weighted Average Layers (Short / Long): Used to calculate the Exponential Moving Average (EMA) according to operator 101 (weighted average), i.e., two layers of EMA short-term and EMA long-term.

[0102] 404 — Subtraction Layer: Calculates the difference between the short-term and long-term EMAs to form the MACD curve; corresponds to operator 102.

[0103] 405 and 406 — Smoothing Layers: Additional weighted averaging layers used to smooth the MACD curve (e.g., over 9 periods) and form signal lines, according to operator 101.

[0104] 407 — Second Subtraction Layer: Calculate the difference between the MACD curve and the signal line to generate a MACD histogram, according to operator 102.

[0105] 408 — Output layer: Outputs the final MACD trading signal.

[0106] The mathematical definition of the MACD indicator:

[0107] MACD t =EMA t 短期 - EMA t 长期

[0108] Signal t =EMA 9 MACD t )

[0109] Histogramm t = MACD t - Signal t

[0110] Mappings in the TIN system: EMA short-term vs. EMA long-term → Operator 101 (layers 402 and 403); MACD = Difference → Operator 102 (layer 404); Signal line = EMA on MACD → Operator 101 (layers 405 and 406); Histogram = Difference between MACD and signal line → Operator 102 (layer 407). This structure completely transforms the mathematical logic of MACD into a reconfigurable neural architecture. The TIN system can initialize the weights of the involved EMA layers using coefficients defined in the classic MACD indicator. Furthermore, these parameters can be adjusted based on data without altering the interpretable structure of the indicator.

[0111] Figure 5 A one-dimensional TIN used for the structured implementation of a technical indicator.

[0112] Figure 5 The detailed implementation of a one-dimensional TIN is shown as an example structure, in which the calculation of a classic indicator (such as MACD or MMA) is completely transformed into a neural network architecture. Figure 1 The defined operators form the basis for transforming mathematical rules into structured layers.

[0113] 501 — Input Layer (Time, Price, or Volume Data): Represents the sequential input values ​​P over N time step windows. t P t−1 、…

[0114] 502 — MA slow layer: according to operator 101 ( Figure 1 Multiple parallel weighted average operations are performed, using weight w. i,s 慢速 Perform slow smoothing.

[0115] 503 — MA Fast Layer: According to operator 101 ( Figure 1 Parallel weighted average operation, using weight w i,k 快速 Calculate the fast average.

[0116] 504 — Subtraction layer: For each pair (according to operator 102), the difference is taken between MA fast and MA slow, thereby generating an initial differential signal.

[0117] 505 — Aggregation / fusion layer: Integrates differential signals, for example, by using weighted sum or logic connections.

[0118] 506 — Terminal Differential Layer: The final output layer of the aggregated subtraction (again, operator 102) is used to generate a tradable signal as output.

[0119] This structure demonstrates how the original mathematical components of technical indicators are transformed into layer operations: MA structure → Operator 101 (weighted sum); Difference → Operator 102 (subtraction); Structured fusion and smoothing → Combination of operators 101 and 102. The network is entirely composed of standardized, reusable components. The original computational logic is preserved but modularly mapped through the TIN system, making it learnable and combinable with other indicator networks or data sources. Therefore, Figure 5 As Figure 3 The illustrated topology is an exemplary refinement, specifically for one-dimensional application scenarios.

[0120] Figure 6 : Scalable TIN (based on MACD) using two-dimensional input data

[0121] Figure 6 This illustrates a network based on a TIN system that processes two-dimensional input data (such as fundamental data, sentiment values, or textual information). The underlying network topology is based on... Figure 4 The MACD model, and represents Figure 5 A scaled extension of the one-dimensional example. This structure clarifies the implementation of classic indicator logic not only on sequential price data but also on higher-dimensional input data.

[0122] 601 — Two-dimensional data input layer: A matrix-style input data structure consisting of features such as financial indicators, sentiment indicators, corporate parameters, or external event volume.

[0123] 602 — Weighted Feature Extraction: According to Operator 101 ( Figure 1 Multiple parallel weighted operators compute aggregated vector structures (similar to MA) for each data axis. i,j ).

[0124] 603 — Subtraction layer: Calculates the difference across multiple aggregation axes according to operator 102, for example, between contrasting fundamental data or sentiment features.

[0125] 604 — Fusion and Smoothing: Differential signals are combined by an adaptive weighting layer (operator 101) to structurally aggregate different influence dimensions.

[0126] 605 — Output layer with decision vectors: following the subtraction structure of operator 102; optional classification logic in addition to operators 101–106, which produces multidimensional outputs such as decision matrices, risk scores, or dynamic thresholds.

[0127] This architecture clarifies how the TIN system extends the original one-dimensional indicator structure to complex input data. Both the input and output layers can be constructed as multi-dimensional. The underlying indicators (such as the MA differential logic in MACD) are preserved in their mathematical structure and transformed into a scalable multi-channel architecture in the TIN system.

[0128] Figure 7 Network structure of the ROC index in the TIN system

[0129] Figure 7 This illustrates the network topology mapped to a classic ROC within a TIN system. This structure is based on... Figure 1 The weighted average, division, and subtraction operators of the defined functional components.

[0130] 701 — Input Layer: Collects sequential input data, such as price data, over a time window N. The data source may also include other trading-related information, such as trading volume, sentiment, or event vectors.

[0131] 702 and 703 — Aggregation layer: according to operator 101 ( Figure 1 Two weighted average layers are used to calculate aggregate values ​​from different reference points, such as P. t and P t−N .

[0132] 704 — Division Level: According to Figure 1 Operator 104 is used to calculate the change quotient between two aggregates.

[0133] 705 and 706 — Comparison Aggregation: Used to calculate smoothed comparison values ​​(e.g., normalization, signal cleanup) according to another averaging layer of operator 101.

[0134] 707 — Subtraction layer: according to operator 102 ( Figure 1 The difference operator is used to determine the final difference signal between the division result and the comparison value.

[0135] 708 — Output Layer (ROC Signal): Provides the final output value of the ROC metric, for example, for classification or threshold testing.

[0136] Mathematical definition of ROC indicator:

[0137] ROC t = (P) t - P t−N ) / P t−N = P t / P t−N -1

[0138] Mappings in the TIN system: Time comparison → Aggregation by operator 101 (layers 702 and 703); Rate of change → Division by operator 104 (layer 704); Signal smoothing or normalization → Operator 101 (layers 705 and 706); Differential signal → Operator 102 (layer 707). This network structure transforms the mathematical definition of ROC into a reconfigurable layer topology. The original formulas are preserved and technically implemented in a structured form within the TIN system.

[0139] Figure 8 Network structure of RSI index in TIN system

[0140] Figure 8 This illustrates the network structure for implementing RSI within a TIN system. Classical RSI logic utilizes... Figure 1 The operators in the model are transformed into a structured architecture where the input and output data are not limited to price information.

[0141] 801 — Input Layer: Collects sequential data streams (such as price, volume, text, or event vectors), not limited to the original RSI definition.

[0142] 802 — Aggregation layer: Calculates aggregated values ​​for preprocessing, serving as the basis for limiting and filtering (operator 101).

[0143] 803 and 804 — Upper / lower limit layers: two separate limiting units used to separate the positive and negative components of the time series, according to operator 105.

[0144] 805 and 806 — Moving average layers: The separated components are smoothed using operator 101 (average) and used as input for intensity analysis.

[0145] 807 — Division layer: formed according to the ratio of operator 104:

[0146] RSI t = Average increase t / (Average increase) t + Average decline t )

[0147] 808 — Output layer: Outputs RSI values, for example as continuous scores, binary trading signals, or decision vectors.

[0148] Mathematical definition of RSI:

[0149] RSI t = 100 − 100 / ( 1 + RS t ), where RS t = Average increase t / Average decline t

[0150] Mappings in the TIN system: Averaging of gains / losses → Operator 101 (layers 805 and 806); Separation of positive / negative values ​​→ Operator 105 (layers 803 and 804); Ratio formation → Operator 104 (layer 807). This network structure transforms RSI into a reconfigurable subnetwork where the mathematical definitions are preserved. The TIN system can also incorporate additional input data into this structure and provide outputs in various signal forms.

[0151] Figure 9 Network structure of the stochastic oscillator in the TIN system

[0152] Figure 9 This illustrates the subnetwork topology representing the random oscillation index within a TIN system. Its mathematical process utilizes… Figure 1 The operators in the network are mapped and transformed into a structured network format.

[0153] 901 — Input Layer: Receives data sources such as prices, highest and lowest values, or external time series data, for example, P. t P high P low The data may also include other transaction-related features, such as fundamental data or the amount of external influence.

[0154] 902 and 903 — Minimum / Maximum Layers: Local extrema are calculated according to operator 106 to determine P over a time window. low and P high .

[0155] 904 and 905 — Subtraction layer: P is calculated according to operator 102 high - P low .

[0156] 906 — Division layer: Calculate the ratio between the positional spacing and the total amplitude according to operator 104:

[0157] (P) t - P low ) / ( P high - P low )

[0158] 907 — Smoothing layer: The %D value is generated by moving average of the %K value according to the average layer of operator 101.

[0159] 908 — Output layer: Calculates random oscillation index signals, such as %K, %D, or structures derived therefrom.

[0160] Mathematical definition:

[0161] %K t = (P) t - P low ) / ( P high - P low ), %D t = MA n (%K) t )

[0162] Mappings in the TIN system: Local extrema → Operator 106 (layers 902 and 903); Subtraction → Operator 102 (layers 904 and 905); Normalization → Operator 104 (layer 906); Smoothing → Operator 101 (layer 907). This network structure transforms stochastic oscillation indices into reconfigurable layer topologies. The original mathematical definitions are preserved and technically implemented in a structured form within the TIN system.

[0163] Figure 10 Network structure of CCI index in TIN system

[0164] Figure 10 The network structure used to map the Commodity Channel Index (CCI) is shown. In this process, the classical CCI calculation steps are transformed into modular network layers.

[0165] 1001 — Input Layer: Receives multidimensional input data (such as highest price, lowest price, and closing price, derivative data, or event data). The data source may also include other transaction-related variables.

[0166] 1002, 1003, and 1004 — Average Value Layer: According to operator 101 ( Figure 1 The three moving average layers represent the typical price TP = (H+L+C) / 3 and its moving average MA. TP .

[0167] 1005 — MAD Layer: Calculates the mean absolute deviation (MAD) of TP relative to its MA according to operator 103. This module directly corresponds to the classical definition.

[0168] 1006 — Subtraction layer: Calculate TP − MA according to operator 102 TP That is, the distance relative to the center of movement.

[0169] 1007 — Division layer: Calculate the normalized CCI value according to operator 104:

[0170] CCI t = (TP) t - MA TP ) / ( 0.015·MAD t )

[0171] 1008 — Output layer: Outputs CCI values ​​or forwards them as input to downstream modules, such as decision logic, classification, or threshold evaluation.

[0172] The classic mathematical definition of CCI:

[0173] CCI t = (TP) t - MA TP ) / ( 0.015·MAD t ), TP t = ( H t + L t + C t ) / 3

[0174] Mapping in the TIN system: TP formation and smoothing → Operator 101 (layers 1002, 1003, and 1004); Bias calculation → Operator 102 (layer 1006); MAD calculation → Operator 103 (layer 1005); Normalization → Operator 104 (layer 1007). This structure allows for a fully reconfigurable implementation of CCI in neural systems and can be extended for multiple data channels, dynamic weights, and flexible decision logic. Layers 1004 to 1008 directly represent the original mathematical structure of CCI without altering the topology, while laying the foundation for complex, context-adaptive applications.

[0175] Example

[0176] The following six TIN examples are described in the form of illustrations, each based on a classic technical analysis indicator.

[0177] Example 1: TIN-MA (Moving Average) (see...) Figure 2 )

[0178] Network Structure: An input layer receives time-sorted price data, trading volume, order book data, sentiment vectors, fundamental data, event information, or other relevant input data from various sources; a sliding aggregation layer performs a weighted average over a defined time window, with its weighting logic based on fixed or dynamically adjustable parameters; a comparison module compares the aggregated average with the current or shifted market price, thereby deriving a directional signal; this component can be implemented as a differential layer or a logical decision layer; optionally, a time-shifted comparison structure is used to identify moving average crossovers.

[0179] Example 2: TIN-MACD (Moving Average Convergence Divergence) (see...) Figure 4 )

[0180] Network structure: an input layer as described above; two parallel aggregation paths with weighted processing units for calculating a "fast" and a "slow" moving average; a subtraction layer for determining the difference between the two moving averages, which generates the raw MACD signal; a post-smoothing layer for processing the difference signal; and an output layer for generating a categorized trading signal based on threshold rules, difference direction, or cross-identification.

[0181] Example 3: TIN-RSI (Relative Strength Index) (see...) Figure 8 )

[0182] Network structure: an input layer as described above; two separate aggregation modules for averaging positive and negative price changes, filtered by a ReLU-like separation function; a division unit with bias regularization for forming the ratio of the averages; a normalization layer for transforming the results to a normalized range of 0 to 100; and an output layer for classifying the RSI levels to a threshold range.

[0183] Example 4: TIN-ROC (rate of change) (see...) Figure 7 )

[0184] Network structure: an input layer as described above; a subtraction module for determining the difference between the current price value and the lagged price value at fixed time intervals; a division unit for normalizing the difference relative to the lagged price value and supplementing it with regularization; a post-scaling unit for representing the relative change value as a percentage; and an output layer for threshold classification.

[0185] Example 5: TIN-CCI (Commodity Channel Index) (see...) Figure 10 )

[0186] Network structure: An input layer as described above; a computation layer for determining the typical price for each period (arithmetic average of the highest, lowest, and closing prices); an aggregation and averaging layer for calculating the moving average of the typical price; a post-MAD layer for determining the average deviation relative to the moving average; a division unit for normalizing the deviation and supplementing it with regularization; and an output layer for classifying CCI values ​​for extreme value regions.

[0187] Example 6: TIN - Stochastic Oscillator (see...) Figure 9 )

[0188] Network structure: An input layer as described above; two minimum / maximum modules for determining local highs and lows within a time window; a difference forming layer for determining the amplitude and the distance from the minimum; a division unit for normalizing the current price level within the amplitude (Stochastic Oscillator %K); a smoothing layer for calculating the Stochastic Oscillator %D using a moving average of the %K values; and an output layer for interpreting the %D signal based on a predefined threshold.

[0189] Industrial applications

[0190] This invention is industrially applicable because it can be transformed into standardized software modules and directly used in operational trading systems. The proposed Technical Indicator Network (TIN system) has a deterministic structure capable of generating technical trading signals based on existing financial data streams. It can implement reconstructed classical technical indicators in the form of modular neural networks, whose weights can be analytically initialized or adaptively trained. This system is specifically designed for the following industrial applications:

[0191] 1. High-Frequency Automated Trading Systems (HFT): TIN systems, due to their deterministic structure and low computational load, are suitable for real-time environments with critical latency requirements. In-system initialization using a fixed weight matrix significantly reduces training complexity, allowing the system to run on FPGA, GPU, or ASIC-based execution units.

[0192] 2. Algorithmic Strategy Generation in Institutional Trading Systems: This system can be integrated into commercial platforms for developing systematic strategies, generating automated decision recommendations based on technical analysis. By modularly embedding multiple Indicator Networks (INs), the TIN system is suitable for multi-asset strategies within institutional portfolios.

[0193] 3. AI Signal Processing for Brokers and Trading Platforms: This system can be integrated into existing brokerage APIs, quantitative research platforms, or trading infrastructures, for example, as a neural backend service for generating buy / sell / hold signals in real time. The trainable parameter structure thus allows for dynamic adaptation to fluctuating market conditions without losing interpretable structure.

[0194] 4. Simulation and Backtesting in Quantitative Analysis Environments: The structure-fidelity mapping of classic indicators in neural networks allows the system to be used for simulation and research systems to evaluate trading strategies. This also enables its industrial applications in risk analysis, performance optimization, and strategy diversification.

[0195] 5. Application in regulated trading systems: Through an interpretable architecture and a clear topological correspondence of mathematical indicator logic, the system can meet the regulatory requirements in the context of "explainable artificial intelligence (XAI)" in the financial industry.

[0196] Implementation typically takes place in standardized programming environments (such as the Python ecosystem with deep learning frameworks like TensorFlow, PyTorch, or ONNX compatible models) and is fully automated. Compatibility with commonly used financial data formats (such as open-high-low-close-volume (OHLCV) tables, order book data structures, or sentiment vectors generated by NLP systems) ensures seamless integration into existing industry workflows.

Claims

1. A system that reconstructs classic technical analysis indicators into a modular neural network, characterized in that, include: The input layer is used to collect and preprocess price data as well as structured and unstructured supplementary data from a variety of different sources, especially those related to financial markets, including but not limited to trading platform API data streams, news feeds, order book data, quantitative sentiment vectors, or other structured market data. At least one index network (IN) whose layer structure maps the topological reconstruction of the mathematical logic of technical indices; One or more processing units serving as neural computation layers, the processing units being used as shown in Figure 1 for aggregation, difference, normalization, amplitude limiting, minimum / maximum determination, division, and transformed vector operations, which correspond to the classical operators of technical indicators; The output layer is used to generate one or more technical output values ​​for transactions, which can be categorized into numerically representable decisions.

2. The system according to claim 1, characterized in that, The indicator network has a sliding aggregation structure based on time series operations and is applicable to moving average (MA) type methods.

3. The system according to claim 1, characterized in that, The index network includes differential operations between aggregation branches and subsequent smoothing via a weighted filtering structure.

4. The system according to claim 1, characterized in that, It integrates a division unit with bias regularization to avoid numerical instability when the denominator is small.

5. The system according to claim 1, characterized in that, The weight parameters of the network layer are determined either by structural assignment mapping based on the original definition of the index, or adaptively through training methods such as reinforcement learning, supervised learning, or other machine learning methods.

6. The system according to claim 1, characterized in that, Multiple indicator networks can be used in parallel, and their outputs can be embedded into a higher-level decision-making or analysis system. This embedding can be performed in an aggregated, adaptive, or sequential manner, including weighted fusion, voting mechanisms, or meta-network-based selection methods.

7. The system according to claim 1, characterized in that, In addition to price data, the input layer can also process structured market data or unstructured text data from news feeds, trading platform API data streams, trading volume, or quantitative sentiment vectors, but is not limited to these examples.

8. A training system within a TIN system for deriving technical trading strategies by selectively modulating weight parameters in a neural network initialized by technical indicator logic, characterized in that... include: An initialization unit, which is applied to the reconstructed topology network layer of the technical indicators mapped within the TIN system; The training module employs reinforcement learning, supervised learning, or other machine learning methods. The training module operates only within the structural constraints of the TIN system and focuses on the technical adjustment of the weights within the indicator structure. The simulated market interface consists of feedback components for transaction costs, position history, and state vectors, and is associated with the TIN topology; The evaluation system is used to derive structured decision quantities, which are directly coupled to the index output of the corresponding TIN; The weight components, whose parameters are determined either by structural assignment mapping based on the original definition of the corresponding technical indicators or by learning-based methods, are used for runtime-related correlation assessment. The result fusion unit is used for signal combination and decision preparation.

Images