CN-121997027-A - AI quantization time sequence characteristic self-adaptive extraction algorithm system

CN121997027ACN 121997027 ACN121997027 ACN 121997027ACN-121997027-A

Abstract

The invention relates to the technical field of artificial intelligence, big data mining and complex time sequence analysis, in particular to an AI quantization time sequence characteristic self-adaptive extraction algorithm system, which comprises a space-time multidimensional self-adaptive purification module, a time-space multidimensional self-adaptive extraction module and a time-space analysis module, wherein the space-time multidimensional self-adaptive purification module is used for optimizing the data signal-to-noise ratio through a fluctuation rate and a time period mask; the system comprises a microstructure perception feature construction module, a dynamic asymmetric fluctuation rate tag generation module, a two-way parallel integrated characterization learning module and a closed loop feedback optimization module, wherein the microstructure perception feature construction module is used for constructing a relative feature space and extracting fractal and entropy features, the dynamic asymmetric fluctuation rate tag generation module is used for generating an equalization sample tag based on average real amplitude, the two-way parallel integrated characterization learning module is used for predicting through fusing a Bagging and Boosting channel, and the closed loop feedback optimization module is used for optimizing system parameters according to simulation objective function feedback. The method comprises the corresponding steps. The method solves the problems of lack of state perception of the features, physical meaning damage caused by sample construction and the like, and realizes time sequence feature extraction and trend identification of high robustness and self-adaptive evolution.

Inventors

DENG XIONG
LI JIACHENG
SHI RUI

Assignees

杭州翊锋量化科技有限公司

Dates

Publication Date: 20260508
Application Date: 20260128

Claims (6)

1. An AI quantization timing characteristic adaptive extraction algorithm system, comprising: The space-time multidimensional self-adaptive purification module is used for carrying out dynamic signal-to-noise ratio optimization on the original time sequence data based on a fluctuation rate system identification and a time domain structure mask mechanism, and eliminating invalid space-time samples; The microstructure perception feature construction module is connected with the space-time multidimensional self-adaptive purification module and is used for constructing a relative feature space of the statistical distribution of the current data point relative to the historical time period based on the market time period state machine and extracting fractal dimension and entropy value features; The dynamic asymmetric fluctuation rate label generation module is connected with the space-time multidimensional self-adaptive purification module and the microstructure perception feature construction module and is used for constructing a dynamic potential energy barrier based on average real wave amplitude and realizing natural equalization of sample types on the premise of not carrying out artificial synthesis sampling through self-adaptive adjustment of asymmetric parameters; The two-way parallel integrated characterization learning module is connected with the microstructure perception feature construction module and the dynamic asymmetric fluctuation rate tag generation module, comprises a Bagging characterization path and a Boosting characterization path which are deployed in parallel, and is used for carrying out high-dimensional mapping and probability fusion on the extracted features; And the closed loop feedback optimization module is connected with the output end of the two-way parallel integrated representation learning module, and is used for feedback-connecting the space-time multidimensional self-adaptive purification module with the parameter input end of the dynamic asymmetric fluctuation rate label generation module, and is used for reversely iteratively optimizing the filtering threshold value of the space-time multidimensional self-adaptive purification module and the asymmetric parameter of the dynamic asymmetric fluctuation rate label generation module based on the simulation cost function of the downstream task.
2. The system of claim 1, wherein the time-space multidimensional adaptive purification module implements a time domain structure mask through a time domain authority mask table, the mask table divides a time axis into a plurality of discrete time periods according to periodic microscopic characteristics of a time sequence, and assigns independent feature extraction authority weights for each time period, and a data party of a corresponding time period can enter a downstream module only when the weights meet preset conditions.
3. The adaptive extraction algorithm of AI quantization timing characteristics as set forth in claim 1, wherein the relative characteristics extracted by the microstructure-aware feature construction module include at least a normalized deviation of the current value from an extremum within a window of a previous time period, a penetration of the current value into a window boundary of the previous time period, and a quantile position of the current value in a window distribution of the previous time period.
4. The AI quantization timing feature adaptive extraction algorithm system of claim 1, wherein the dynamic asymmetric volatility tag generation module is specifically configured to: Based on the current average real wave amplitude, multiplying the current average real wave amplitude by an independently adjustable positive threshold coefficient and a negative threshold coefficient respectively to form an asymmetric dynamic classification boundary; and automatically searching the optimal positive threshold coefficient and the optimal negative threshold coefficient through the closed loop feedback optimization module, so that the generated labels meet the maximum entropy constraint on the statistical distribution.
5. The adaptive extraction algorithm system of AI quantization timing characteristics according to claim 1, wherein the objective function constructed by the closed-loop feedback optimization module is configured to forcibly subtract the simulated transaction friction cost and the time value loss when evaluating the performance of the model, and dynamically adjust the effective interval threshold of the fluctuation rate in the space-time multidimensional adaptive purification module according to the evaluation result based on the objective function.
6. An AI quantization timing characteristic self-adaptive extraction method is characterized by comprising the following steps: Performing space-time multidimensional self-adaptive purification on the original time sequence data, and removing invalid samples based on the fluctuation rate system identification and the time domain structure mask mechanism; based on a market period state machine, constructing a relative feature space for the purified data, and extracting fractal dimension and entropy value features; Dynamically generating a label for realizing natural equalization of sample types based on the average real amplitude and the asymmetric parameters; utilizing a Bagging path and a Boosting path which are deployed in parallel to perform integrated characterization learning and probability fusion on the characteristics; And performing closed-loop feedback optimization on the filtering threshold value of the purification process and the asymmetric parameters generated by the labels based on the simulation cost function of the downstream task.

Description

AI quantization time sequence characteristic self-adaptive extraction algorithm system Technical Field The invention relates to the technical field of artificial intelligence, big data mining and complex time sequence analysis, in particular to an AI quantization time sequence characteristic self-adaptive extraction algorithm system. Background In the field of complex time sequence analysis, in particular to foreign exchange finance quantization analysis, how to extract the characteristics with high generalization capability from massive, high-noise and nonlinear original data is a key for constructing a high-performance prediction model. The prior art scheme mainly faces the following deep technical bottlenecks: The feature extraction lacks 'state perception', and the traditional technology adopts a sliding window to calculate statistical indexes (such as mean line and RSI). However, the financial market has obvious microstructural differences (e.g., different fluidity profiles for different time periods, different noise levels for different volatility intervals). The traditional method adopts a uniform operator on all time slices, ignores the 'context state' of the data, and leads to the failure of the features under different market systems. The label construction and sample equalization 'pseudo proposition' is to solve the problem of unbalance of positive and negative samples, and the prior art usually adopts artificial synthesis sampling (such as SMOTE) or simple oversampling. The method breaks the continuity and microscopic correlation of the time sequence in a physical sense, and introduces false high-frequency noise. Meanwhile, the tag generation method based on a fixed threshold (such as fixed fluctuation amplitude) cannot adapt to fluctuation rate changes of different assets in different periods, so that the signal-to-noise ratio of the tag is extremely unstable. The model complexity is not matched with the data entropy value, namely along with the development of deep learning, transformer, CNN-LSTM and other high-complexity models are widely applied. However, in financial timing with low signal-to-noise ratio, excessive model complexity tends to capture random noise rather than deterministic rules. The lack of a closed-loop feature evaluation mechanism is that the existing feature engineering is often unidirectional, namely 'extracting features- > inputting a model', and the lack of a feedback loop based on terminal tasks (such as actual predicted benefits and risk adjusted benefits) automatically optimizes the super-parameters of feature extraction. Therefore, an algorithm system capable of adaptively sensing the microscopic state of the market, dynamically adjusting the sample construction logic and continuously optimizing the feature expression capacity through closed-loop feedback is developed, and the algorithm system has important theoretical significance and application value. Disclosure of Invention The invention aims to provide an AI quantization time sequence characteristic self-adaptive extraction algorithm system so as to solve the problems in the background technology. In order to solve the technical problems, the invention provides the following technical scheme: in a first aspect, the present invention provides an AI quantization timing feature adaptive extraction algorithm system, including: The space-time multidimensional self-adaptive purification module is used for carrying out dynamic signal-to-noise ratio optimization on the original time sequence data based on a fluctuation rate system identification and a time domain structure mask mechanism, and eliminating invalid space-time samples; The microstructure perception feature construction module is connected with the space-time multidimensional self-adaptive purification module and is used for constructing a relative feature space of the statistical distribution of the current data point relative to the historical time period based on the market time period state machine and extracting fractal dimension and entropy value features; The dynamic asymmetric fluctuation rate label generation module is connected with the space-time multidimensional self-adaptive purification module and the microstructure perception feature construction module and is used for constructing a dynamic potential energy barrier based on average real wave amplitude and realizing natural equalization of sample types on the premise of not carrying out artificial synthesis sampling through self-adaptive adjustment of asymmetric parameters; The two-way parallel integrated characterization learning module is connected with the microstructure perception feature construction module and the dynamic asymmetric fluctuation rate tag generation module, comprises a Bagging characterization path and a Boosting characterization path which are deployed in parallel, and is used for carrying out high-dimensional mapping and probability fusion on the extracted features; And