A financial market fluctuation prediction system based on time series pattern mining

Abstract

The application discloses a financial market fluctuation prediction system based on time sequence pattern mining, and relates to the technical field of financial information.The system comprises a pattern deconstruction unit, a relationship network construction unit, a real-time sensing and deduction unit and a state prediction unit.The pattern deconstruction unit is used for dividing a historical fluctuation rate sequence into macro state intervals and extracting fine-grained basic fluctuation pattern sequences of the intervals.The relationship network construction unit is used for constructing a pattern relationship network comprising pattern transition probability, symbiotic strength and state transition probability according to historical data.The real-time sensing and deduction unit is used for identifying real-time patterns and deducing future multiple possible basic pattern paths based on the relationship network.The state prediction unit is used for matching and reconstructing micro pattern paths into a macro state sequence and outputting a prediction probability distribution of a future state.The application realizes the prediction of market fluctuation macro states, improves the early warning capability of market state turning points and the interpretability of prediction results, and can be directly mapped into dynamic risk parameters, thereby providing effective decision support for risk management.

Description

Technical Field

[0001] This invention relates to the field of financial information technology, and in particular to a financial market volatility prediction system based on time series pattern mining. Background Technology

[0002] Accurate prediction of financial market volatility is crucial for risk management, asset pricing, and investment decisions. Traditional volatility prediction methods, such as the Generalized Autoregressive Conditional Heteroskedasticity (GARCH) family of models and their extensions, primarily rely on parametric modeling of the statistical properties of volatility sequences themselves. In recent years, with the development of machine learning techniques, deep learning models such as Long Short-Term Memory (LSTM) networks, Gated Recurrent Units (GRUs), and Transformers have been widely applied to time series forecasting tasks, attempting to learn complex nonlinear mapping relationships from historical data.

[0003] However, the aforementioned existing technical solutions have the following limitations: First, both traditional econometric models and deep learning models primarily focus on directly fitting and regressing volatility values ​​for prediction. These methods struggle to explicitly capture and utilize multi-scale time-series patterns with clear economic or behavioral finance implications within market fluctuations (e.g., volatility clustering, state transitions, and other micro and macroeconomic phenomena). Second, these methods, especially complex deep learning models, are often considered "black boxes," lacking interpretability in their internal decision-making mechanisms. This makes it difficult for practitioners to understand and trust the prediction results, thus limiting their application in critical decision-making. Finally, the output of existing methods is typically a single value or a narrow range estimate of future volatility. This form of prediction makes it difficult to directly and effectively automate and refine downstream risk control rules (such as dynamic position adjustments and stop-loss threshold settings).

[0004] Therefore, designing a forecasting system that can uncover the inherent patterns in volatility time series, provide interpretable state predictions, and directly support risk management decisions is a technical problem that urgently needs to be solved. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a financial market volatility prediction system based on time series pattern mining to address the problems raised in the background art. Specifically, existing financial market volatility prediction methods mostly focus on the direct fitting and extrapolation of historical numerical sequences, failing to fully explore and utilize the multi-scale time series patterns and their complex transition and symbiotic relationships inherent in the volatility evolution process. This results in insufficient early warning of key turning points in market volatility states (or "phases") and weak interpretability of the prediction models.

[0006] To address the aforementioned technical problems, this invention provides a financial market volatility prediction system based on time-series pattern mining. This system predicts macroeconomic market volatility by performing hierarchical pattern deconstruction on volatility sequences, constructing a pattern relationship network, and dynamically extrapolating based on this network.

[0007] According to one aspect of the present invention, a financial market volatility prediction system based on time series pattern mining is provided, comprising: The pattern deconstruction unit is used to process historical volatility time series data. It performs the following actions: identifying statistical feature abrupt changes in the sequence to divide the macro state intervals; performing morphological analysis on the data segments within each interval, and classifying them into a finite number of basic volatility patterns according to predefined rules, thereby obtaining the basic pattern sequence corresponding to each macro state interval.

[0008] A relational network construction unit, connected to the pattern deconstruction unit, is used to construct a pattern relational network based on historical data. The network stores the transition probabilities between basic patterns, the symbiotic association strength between basic patterns, and the transition probabilities between macroscopic states.

[0009] The real-time sensing and inference unit, connected to the relational network construction unit, is used to process real-time volatility data. It performs the following actions: identifying the basic volatility pattern of the current data segment and updating the current pattern sequence; starting from the current pattern sequence, using the transition probability and symbiotic association strength in the pattern relational network to perform probabilistic inference, generating multiple future basic pattern evolution paths and probability weights for each path.

[0010] The state prediction unit, connected to the real-time perception and inference unit and the relationship network construction unit, is used to map each future evolution path into a macroscopic state sequence. It performs the following steps: dividing the path into segments; matching the pattern set of each segment with the pattern set of the historical macroscopic state interval to determine the predicted macroscopic state corresponding to the segment; and combining the matching results and probability weights of all paths to output the probability distribution of the macroscopic state in each future time period.

[0011] As a further optimization, the pattern deconstruction unit identifies abrupt change points by comparing whether the differences in statistics (such as mean and variance) between sliding windows exceed a dynamic threshold determined based on historical data.

[0012] As a further optimization, the basic fluctuation pattern is defined by analyzing the sign combination of the first-order difference mean and the second-order difference mean of the data segment, such as accelerating upward, decelerating upward, accelerating downward, and decelerating downward.

[0013] As a further optimization, the co-occurrence strength in the pattern relationship network is quantified by statistically analyzing the pattern co-occurrence frequency within a set time window and calculating the point mutual information value.

[0014] As a further optimization, the real-time perception and inference unit uses a beam search algorithm for inference, where the probability of path expansion is obtained by adjusting the basic transition probability with the strength of co-occurrence association.

[0015] As a further optimization, the state prediction unit measures the matching degree between the fragment pattern set and the historical macroscopic state pattern set by calculating the Jaccard similarity coefficient.

[0016] As a further optimization, the system also includes a risk parameter mapping unit, which is used to calculate and output dynamic risk parameter suggested values ​​based on the macroscopic state probability distribution and the preset state-risk parameter mapping relationship.

[0017] As a further optimization, the system also includes an online update unit for triggering updates to the pattern relationship network based on prediction accuracy or a fixed period.

[0018] Compared with the prior art, the technical solution provided by the present invention can achieve the following beneficial effects: Compared to existing technologies, this system continuously analyzes the micro-evolutionary sequence of "basic fluctuation patterns" in real-time data and infers their development trends based on pattern relationship networks. This process can identify early changes in the microstructure before significant anomalies occur in macro-statistical indicators reflecting the overall market level, thus providing earlier warnings of potential turning points in the "macro-state."

[0019] Compared to existing technologies, the final output of this system is a probabilistic prediction of a "macroeconomic state" (such as high volatility oscillations). This result directly corresponds to a clearly defined and understandable market phase in historical data, making the prediction conclusion intuitive and clear. Decision-makers can directly invoke historical experience or preset rules associated with this state based on this state information, thereby improving the pertinence and efficiency of risk decision-making.

[0020] Compared with existing technologies, this system automatically generates dynamic and continuous risk control parameters (such as position size and stop-loss line) based on the probability distribution of macroscopic states, through a preset state-parameter mapping table and weighted average calculation. This method avoids the execution risk of parameter mutations in traditional binary decision-making, and achieves refined and reliable adjustment of risk control.

[0021] Compared to existing technologies, this system can update its core pattern relation network library using new data based on predicted performance or a fixed period. This mechanism enables the system to proactively track and integrate new features and patterns that emerge during market evolution, thereby effectively mitigating the problem of predictive power decay caused by market "concept drift" in static models.

[0022] Compared to existing technologies, this system completes the complex pattern mining and relation learning processes offline, while the online prediction stage primarily involves rapid inference and matching based on lightweight probabilistic networks. This architecture enables the system to perform deep pattern reasoning without running complex computational models in real time, meeting the high real-time requirements of financial applications. Attached Figure Description

[0023] Figure 1 This is a flowchart of the overall system processing of the present invention.

[0024] Figure 2 This is a flowchart of the pattern deconstruction and network construction process of the present invention.

[0025] Figure 3 This is a flowchart of the real-time simulation and state prediction branch of the present invention. Detailed Implementation

[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] Example 1 As attached Figures 1 to 3 The financial market volatility prediction system based on time-series pattern mining is shown below. The specific implementation content includes three main processes: historical data parsing and knowledge base construction, real-time data prediction, and system maintenance.

[0028] S100: Historical data parsing and pattern relation network library construction.

[0029] This stage extracts multi-level patterns based on historical market volatility time-series data and constructs their relationship network to form a systematic pattern relationship network library.

[0030] S101: Obtain historical volatility time series and perform preprocessing.

[0031] Obtain historical financial market volatility time-series data, which is a set of values ​​arranged in chronological order, denoted as a sequence. ,in The total number of data points. Indicates the first Volatility per unit of time.

[0032] For sequence Standardization is performed, specifically by calculating the mean of the sequence. and standard deviation Then transform each data point. The mean was obtained as Standard deviation is new sequence of .

[0033] S102: Divide the macroeconomic fluctuation state interval.

[0034] Preprocessed sequences The analysis was conducted to identify the points in time when the statistical characteristics changed significantly, and the continuous sequence was divided into multiple macroscopic state intervals accordingly.

[0035] A change-point detection technique is employed. Specifically, a statistical test method based on a sliding window is used. A window with a length of... A sliding window, the length of which is... For example, for 30 data points, the value range can be from 20 to 60 data points.

[0036] Slide the window along the sequence, for each window position ( ), calculation window mean of internal data and standard deviation Calculate the difference in statistics between adjacent windows and .

[0037] Based on all historical data and For each sequence, the 90th percentile is calculated as a dynamic threshold. and .when or At that time, determine the time point To count mutation points. All mutation points divide the sequence into... A continuous interval.

[0038] Each interval is defined as a macro-level fluctuation state interval and assigned a macro-level state identifier. ( The macroeconomic state identifier can be named based on a combination of the average level (high, medium, low) and the trend (rising, falling, oscillating) of volatility within the range.

[0039] S103: Extract the basic fluctuation pattern sequence.

[0040] For each macroscopic state interval defined in S102 We conduct fine-grained morphological analysis on its internal volatility changes, representing it as a combination of a series of basic volatility patterns.

[0041] Using a fixed length The analysis window slides across the data within the interval, with a sliding step size of [missing value]. Data points. Length of the analysis window. For example, with 10 data points, the range of values ​​could be: Up to 20 data points. For each extracted length is... subsequence window By analyzing the first and second-order characteristics of its numerical changes, it is classified as a predefined basic fluctuation pattern.

[0042] Define four basic patterns: (Accelerated upward trend) (Deceleration-upward type) (Accelerated descent type) (Decelerating descent type), classification rules are based on the first-order difference mean of the window data. and the second difference mean A combination of symbols.

[0043] Calculate the first-order difference sequence its first difference mean .

[0044] Calculate the second-order difference sequence Its second-order difference mean .

[0045] Set a positive threshold for determining the sign. (For example, This is to avoid interference from numerical noise.

[0046] The classification rules are as follows: like and , then model; like and If so, it is in P_B mode; like and , then model; like and , then model.

[0047] The absolute value of the first-order difference mean or the second-order difference mean is not greater than the threshold. In such cases, its sign can be considered as zero, or it can be classified into one of the four pattern categories mentioned above that is closest to the current numerical sign.

[0048] Traversing a macroscopic state interval After analyzing all windows, the underlying volatility pattern sequence corresponding to that interval is obtained, denoted as... ,in , This represents the number of patterns contained in this interval.

[0049] S104: Construct a pattern relation network library.

[0050] The set of macroscopic state intervals obtained based on S102 and the basic wave pattern sequence obtained from S103 Build a schema relation network library to store the following three types of relationship data: 1. Basic Pattern Transition Probability Table ( ): Describes the temporal transition relationship between basic fluctuation patterns.

[0051] Statistical analysis of all historical fundamental volatility pattern sequences In any mode The pattern will appear immediately afterwards. frequency Calculate the transition probability ,in Traverse all basic pattern types .Will Store as a matrix The elements, where the row index corresponds to the predecessor pattern. The column index corresponds to the successor schema. .

[0052] 2. Basic Pattern Coexistence Strength Table ( ): Quantifies the correlation strength when different basic patterns co-occur in a nearby time period.

[0053] Define a co-occurrence detection window span The span For example, Each pattern index position can have a value range of 1. Up to 10 pattern index positions. Traverse each pattern sequence. For each position in the sequence Inspection window The set of patterns within (deduplicated).

[0054] Count the number of times any two different patterns P_u and P_v appear in the same window. Simultaneous statistical model Total number of times it appears in any window and total number of windows Calculate point mutual information: ,Will Store as a Symmetric matrix Element.

[0055] 3. Macroscopic state transition probability table ( ): Describes the transition relationship between macroscopic states.

[0056] Based on the order in which historical macro-state intervals appeared, statistics were compiled from the state... Transferred to frequency Calculate the transition probability ,in Iterate through all macroscopic states, and Stored as a matrix Element.

[0057] S200: Real-time data stream processing and fluctuation state prediction.

[0058] This phase operates online, processes real-time inflow volatility data, and performs multi-step extrapolation based on the pattern relationship network library built in the S100 phase, outputting predictions of future macroeconomic volatility.

[0059] S201: Real-time data caching and pattern awareness.

[0060] Continuously receive real-time volatility data streams Maintain a length of Data cache area ( (Consistent with the definition in S103). Whenever new data arrives, update the cache to ensure it always contains the latest information. Data points Using the same pattern definition and classification rules as in S103, for The analysis yields the underlying volatility pattern at the current moment. Save recent A historical pattern (e.g., ), forming the current mode context sequence .

[0061] S202: Probabilistic projection of future basic model paths.

[0062] With current mode context sequence As the initial state, utilize the pattern relation network library and This involves extrapolating multiple basic pattern paths and their probabilities that may occur at various future time steps.

[0063] A beam search algorithm is used for deduction. The beam width is set. (For example, ) and deduction step size (For example, ), initialize a path set The set contains a path whose sequence is: And set its initial path probability. .

[0064] For the deduction step from arrive Execute iteration: (1) Create an empty list of candidate paths .

[0065] (2) For the current path bundle Each path in : Take the last pattern of the path. .

[0066] Query Find out from The first one with the highest probability of departure A set of candidate successor patterns (For example, ).

[0067] For each candidate successor pattern : Calculate the transition probability .

[0068] Calculate the co-occurrence adjustment factor: extract the last element of the current path sequence. A pattern (e.g., ), query Calculate these patterns and Average coexistence strength Adjustment factor ,in For adjustment coefficients (e.g., ).

[0069] Calculate the probability of the new path .

[0070] Append the new path (sequence is seq) ) and their probability Add to candidate list .

[0071] (3) All candidate paths according to their probability Sort in descending order.

[0072] (4) Select the top-ranked The candidate paths constitute a new path bundle. .

[0073] go through After one iteration, the path bundle Included Path and its normalized probability weights (satisfy The result is the deduction.

[0074] S203: Macroeconomic fluctuation state prediction generation.

[0075] The set of micro-mode paths generated by S202 is mapped to a prediction of macro-fluctuation states.

[0076] Set the segment length SegLen to the number of basic patterns contained in each historical macro state interval. of the median.

[0077] For each candidate pattern path This is then divided into several consecutive segments of length SegLen. For each segment Seg, the following operations are performed sequentially: (1) Extract the set of basic patterns that appear in the fragment. (Duplicate element removal).

[0078] (2) With all historical macroeconomic state intervals The corresponding set of basic patterns Perform a similarity comparison.

[0079] This is achieved by calculating the Jaccard similarity coefficient: ,in This indicates the number of elements in the set.

[0080] (3) Select the similarity with Seg The highest historical macro state This is used as the prediction state of the segment Seg.

[0081] After processing all segments of a path, the macroscopic state prediction sequence corresponding to that path is obtained. .

[0082] Integrate all The prediction results for each path. For each future prediction period. (Indexed by segment order), the macroscopic state predicted for all paths during this period is statistically analyzed.

[0083] Each status MS during this period Predicted probability The calculation is as follows: Output the probability distribution of macroscopic states for each future time period. .

[0084] S300: Application interface and system adaptive maintenance.

[0085] S301: Dynamic mapping of risk decision parameters.

[0086] Macroeconomic condition forecasts are transformed into risk control parameters.

[0087] A pre-defined mapping table defines the benchmark risk parameter value corresponding to each macroeconomic state identifier (MS), such as the maximum position ratio. and stop loss range .

[0088] Receive future time periods from S203 output Macroscopic state probability distribution For a certain period of time in the future The recommended dynamic value for the position ratio. Calculated using the following formula: ; Perform similar calculations on all risk parameters that need to be mapped to generate a dynamic sequence of risk parameter recommendations.

[0089] S302: Periodic updates to the pattern relation network library.

[0090] To adapt to market evolution, the pattern relationship network database is updated periodically.

[0091] Monitor short-term forecasting performance, such as calculating past... The prediction accuracy Acc over a time period (e.g., The actual observation state is determined by... The macroeconomic state identifier is determined by offline analysis of the actual market volatility data corresponding to each time period, using the same process as S102 and S103. Prediction accuracy (Acc) is defined as the proportion of predicted states that match the actual observed states.

[0092] When the prediction accuracy Acc is lower than a preset threshold (e.g., ), or the time elapsed since the last update reaches a fixed period (e.g., When the time unit reaches (a certain number of time units), the update process is triggered.

[0093] After the update is triggered, the most recent Market data from a specific time period is used as new training samples (e.g., ...). Re-execute all steps of stage S100 (S101 to S104) with this new sample to generate a new pattern relation network library (containing updated...). , and ), and replace the old knowledge base for subsequent real-time predictions.

[0094] Finally, the following points should be noted: First, in the description of this application, it should be noted that, unless otherwise specified and limited, the terms "installation", "connection", and "linkage" should be interpreted broadly, and can be mechanical or electrical connections, or internal connections between two components, or direct connections. "Up", "down", "left", "right", etc. are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may change. Secondly: The accompanying drawings of the embodiments disclosed in this invention only involve the structures involved in the embodiments disclosed in this invention. Other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other. In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A financial market volatility prediction system based on time-series pattern mining, characterized in that, include: The pattern deconstruction unit is used to process historical volatility time-series data and performs the following operations: S1. Identify statistical feature mutation points in the time series data, and divide the time series data into multiple macroscopic state intervals based on the mutation points; S2. For each of the macroscopic state intervals, divide its internal data into continuous data segments, calculate the first-order difference mean and the second-order difference mean of each data segment, and classify the data segment into one of the four predefined basic fluctuation patterns according to the sign combination of the first-order difference mean and the second-order difference mean, thereby representing each macroscopic state interval as a basic pattern sequence composed of the basic fluctuation patterns. A relational network construction unit, connected to the pattern deconstruction unit, is used to construct and store a pattern relational network based on all historical macro-state intervals and their corresponding basic pattern sequences. The pattern relational network includes: The basic pattern transition probability matrix, whose elements represent the probability of a second basic fluctuation pattern occurring after the first basic fluctuation pattern; The co-occurrence strength matrix of basic patterns, whose elements represent the correlation strength of different basic wave patterns co-occurring within a preset time window; The macroscopic state transition probability matrix, whose elements represent the probability of the second macroscopic state interval occurring after the first macroscopic state interval; The real-time sensing and inference unit, connected to the relationship network construction unit, is used to perform the following operations: S3. Receive real-time volatility data, identify the underlying volatility pattern corresponding to the current data window using the same classification method as in step S2, and update the current pattern sequence; S4. Starting from the current mode sequence, perform probability deduction based on the basic mode transition probability matrix and the basic mode co-occurrence strength matrix to generate multiple future basic mode candidate paths and assign probability weights to each path. The state prediction unit, connected to the real-time sensing and inference unit and the relationship network construction unit, is used to perform the following operations: S5. Divide each of the aforementioned future basic pattern candidate paths into multiple consecutive segments; S6. For each segment, match the set of basic fluctuation patterns it contains with the set of basic fluctuation patterns corresponding to each historical macroeconomic state interval, and take the historical macroeconomic state with the highest matching degree as the predicted state of the segment. S7. Summarize the predicted states and probability weights of all candidate paths in each future time period to generate the macroscopic state probability distribution for each future time period.

2. The financial market volatility prediction system based on time-series pattern mining according to claim 1, characterized in that, The step S1 of identifying statistical feature mutation points specifically includes: calculating the mean and standard deviation of the time series data within a sliding window; calculating the difference between the mean and standard deviation between adjacent sliding windows; and determining the window boundary as a mutation point when the difference exceeds a dynamic threshold determined based on the percentile of the historical difference sequence.

3. The financial market volatility prediction system based on time-series pattern mining according to claim 1, characterized in that, The four predefined basic fluctuation patterns are determined based on the positive and negative signs of the first-order difference mean and the second-order difference mean, corresponding to four forms: accelerated rise, decelerated rise, accelerated fall, and decelerated fall, respectively.

4. The financial market volatility prediction system based on time-series pattern mining according to claim 1, characterized in that, The correlation strength in the basic mode co-occurrence strength matrix is ​​characterized by calculating the point mutual information value.

5. A financial market volatility prediction system based on time-series pattern mining according to claim 1, characterized in that, The probability deduction in step S4 is performed using a beam search algorithm, specifically including: selecting candidate successor modes based on the basic mode transition probability matrix, calculating a co-occurrence adjustment factor based on the basic mode co-occurrence strength matrix to weight the transition probabilities, and iteratively generating a specified number of high-probability candidate paths.

6. A financial market volatility prediction system based on time-series pattern mining according to claim 1, characterized in that, The matching in step S6 is achieved by calculating set similarity, where the matching degree is the Jaccard similarity coefficient.

7. A financial market volatility prediction system based on time-series pattern mining according to claim 1, characterized in that, The length of the continuous segment in step S6 is determined based on the median number of basic fluctuation patterns contained in the historical macroeconomic state interval.

8. A financial market volatility prediction system based on time-series pattern mining according to any one of claims 1 to 7, characterized in that, It also includes a risk parameter mapping unit, which is connected to the state prediction unit and is used to perform the following operations: calculate the dynamic risk parameter values ​​for each future period by weighted summation based on a predefined state-risk parameter mapping table and the macroscopic state probability distribution for each future period.

9. A financial market volatility prediction system based on time-series pattern mining according to claim 1, characterized in that, It also includes an online update unit, connected to the relationship network construction unit and the state prediction unit, for performing the following operations: monitoring the prediction accuracy; when the accuracy is lower than a threshold or reaches a fixed update cycle, triggering the pattern deconstruction unit and the relationship network construction unit to update the pattern relationship network with new data.

10. A financial market volatility prediction system based on time-series pattern mining according to claim 9, characterized in that, The conditions for triggering an update by the online update unit include: the prediction accuracy is lower than a first threshold, or the time since the last update reaches a preset period.

Images