Quantitative investment strategy generation system and method based on multi-modal data and adaptive reinforcement learning

Abstract

The application discloses a kind of quantitative investment strategy generation system and method based on multi-modal data and adaptive reinforcement learning.The system includes multi-modal data acquisition fusion module, deep feature extraction module, hierarchical reinforcement learning strategy generation engine, online adaptive learning module and risk constraint transaction execution module.Deep feature extraction module uses improved time convolution network and financial Transformer to extract price and text features, and generates market state representation through cross-modal attention fusion;Hierarchical reinforcement learning engine includes meta-strategy network and multiple execution strategy network, meta-strategy identifies market state and activates corresponding execution strategy network to generate transaction action;Online adaptive learning module realizes strategy continuous evolution through multiple copies parallel exploration and population evolution;Risk constraint module uses differentiable constraint network to correct action linear inequality projection, to ensure that instruction is in compliance.The application realizes multi-modal data fusion, market state adaptation, yield risk joint optimization and strategy automatic evolution, effectively improves the sharp ratio and robustness of strategy, reduces the maximum drawdown and risk control violation rate.

Description

Technical Field

[0001] This invention belongs to the fields of financial technology and artificial intelligence technology.

[0002] Specifically, the present invention relates to a system and method for automatically and intelligently generating and continuously optimizing quantitative investment strategies by combining multi-source heterogeneous data (structured and unstructured), deep feature extraction technology and an adaptive hierarchical reinforcement learning framework. Background Technology

[0003] Brief description of existing technologies: Traditional quantitative investment strategies mainly rely on structured data such as historical prices and trading volumes, and use statistical methods or simple machine learning models (such as linear regression and support vector machines) to discover patterns.

[0004] Defects / deficiencies of existing technology: The data dimension is too narrow: it ignores the alpha signals (excess return prediction information) contained in unstructured data such as news sentiment, social media sentiment, macroeconomic reports, and industrial chain data.

[0005] Poor strategy adaptability: When market styles shift, strategies based on fixed rules are prone to failure, requiring frequent manual parameter adjustments and lacking adaptability.

[0006] Lagging risk control: Traditional risk control is mostly based on static thresholds (such as fixed stop-loss lines), which cannot dynamically adjust asset allocation and risk exposure according to market conditions.

[0007] The strategy generation process is not highly automated: it relies heavily on the subjective experience and trial and error of researchers, resulting in long development cycles and low efficiency.

[0008] Technical requirements: Therefore, there is an urgent need for a system and method that can automatically, continuously, and dynamically generate and optimize quantitative investment strategies, effectively integrate multimodal data, adapt to market changes, and embed dynamic risk control. Summary of the Invention

[0009] This invention aims to overcome the aforementioned shortcomings of the prior art and provide a system and method capable of automatically, continuously, and dynamically generating and optimizing quantitative investment strategies, specifically addressing the following technical problems: Technical issues

[0010] The problem of unified feature representation and dynamic fusion of multimodal heterogeneous financial data: Existing methods cannot effectively integrate structured data (market data, fundamental data) and unstructured data (news, social media, research reports), resulting in low information utilization and loss of forward-looking signals.

[0011] The problem of adaptive strategy switching under sudden market changes: Single strategy networks or fixed rule strategies cannot recognize the switching of market styles (bull market, bear market, sideways market), and their performance drops sharply when the environment changes. They also lack explicit state recognition and strategy switching mechanisms.

[0012] The problem of deep coupling and joint optimization between risk management and strategy decision-making: Traditional post-filtering risk control results in a large number of invalid instructions, the strategy network cannot perceive the risk control boundary, and the profit and risk objectives are separated, making it difficult to achieve end-to-end joint optimization.

[0013] The challenges of automated and continuous evolution of strategies and efficient online learning: Existing solutions rely on manual tuning and backtesting, resulting in long iteration cycles; online learning methods are prone to forgetting historical experience or getting stuck in local optima, and cannot quickly adapt to market model shifts. Technical solution

[0014] System components: Multimodal financial data acquisition and fusion module This data is used to collect and preprocess structured and unstructured data, constructing samples aligned to time. Structured data includes high-frequency / daily market data for stocks / futures (open, high, low, close, and volume) and financial indicators (price-to-earnings ratio, price-to-book ratio, return on equity, etc.); unstructured data includes financial news headlines / text, company announcements, and social media posts.

[0015] Deep feature extraction module It contains three sub-modules: Improved Temporal Convolutional Network (TCN): Used to extract local and global patterns of price series, it includes gated residual connections and hybrid dilated convolutional layers, which can capture multi-scale temporal dependencies without revealing future information.

[0016] Financial Adaptive Transformer Encoder: Used to extract semantic features from text data, employing a local-global sparse attention mechanism and learnable temporal decay positional encoding to enhance the understanding of the impact of financial terms and events.

[0017] Cross-modal attention fusion layer: used to dynamically calculate the weights of features from different modalities, and generate a weighted fused market state representation vector St through learnable attention parameters. St .

[0018] Hierarchical reinforcement learning policy generation engine It includes a meta-policy network and multiple execution policy networks: Meta-policy networks use market state representation vectors St StAs input, output the probability distribution of the current market state label (e.g., "high volatility bull market", "low volatility sideways market", "high volatility bear market"), and activate the corresponding execution strategy network.

[0019] The execution policy network (the activator) receives the same St St Output the original trading action: atraw=[Operation Type, Target Code, Position Size]. at raw=[Operation Type, Target Code, Position Ratio]. The operation type is a discrete value (buy / sell / hold), and the position ratio is a continuous value (0~1).

[0020] Online adaptive learning module This system is used to run multiple policy replicas in parallel in a simulated environment, employing an ε-greedy policy for exploration and exploitation. The meta-controller periodically evaluates each replica based on performance metrics (Sharpe ratio, Kamma ratio, maximum drawdown), discards the bottom 20% of replicas, replicates the parameters of the top 20% of replicas and adds random noise for mutation, and then synchronizes the optimal replica parameters to the main policy network to achieve population-based evolution of the policy.

[0021] Risk-constrained transaction execution module An embedded differentiable constraint network is used to express the risk management rules as a system of linear inequalities Aa≤b. Aa ≤ b , where a a Let A be the action vector. A Let b be the constraint coefficient matrix. b This is the constraint threshold vector. Before the action is output, the network checks whether the original action satisfies the constraint; if it is violated, it solves for min∥a′−a∥2 st Aa′≤bmin∥ using projective gradient descent. a ′− a ∥2 st Aa '≤ b Obtain compliance action atcomp at comp, and outputs the executable trading instructions.

[0022] Methods and Steps A method for generating quantitative investment strategies for the above-mentioned system includes the following steps: S1 Data Acquisition and Fusion: Collect and fuse multimodal financial time-series data, including structured market / fundamental data and unstructured text data, and construct a unified sample by aligning it with time.

[0023] S2 Deep Feature Extraction: Deep features are extracted using an improved TCN and Transformer, and then fused through cross-modal attention to form a market state representation vector St. St .

[0024] S3 hierarchical reinforcement learning decision-making: S3.1 will St St Input the meta-policy network, output the market state label (such as "high volatility bear market"), and activate the corresponding execution policy network; S3.2 The activated execution policy network receives St St Output the original transaction action atraw at raw.

[0025] S4 Risk Constraint Verification and Correction: [This section appears to be incomplete and requires further context.] at The raw data is fed into the risk constraint network to check the linear constraint Aa≤b. Aa ≤ b If a violation occurs, it is corrected to a compliant action via projected gradient descent (atcomp). at comp.

[0026] S5 Execution and Feedback Learning: Perform compliant actions, collect environmental feedback (returns, maximum drawdown, transaction costs), and calculate the reward Rt=rt−λ⋅MDDt−γ⋅TCt Rt = rt - λ ⋅MDD t - γ ⋅TC t This is used to update the policy network parameters.

[0027] S6 Population Evolution: The meta-controller asynchronously evaluates the performance of policy replicas, performs elimination, replication, and mutation operations, and migrates the optimal parameters to the main policy network.

[0028] Key technological innovations Innovation Technical means Technical effect Multimodal fusion Improved TCN (Gated Residual + Hybrid Dilated Convolution) + Financial Transformer (Sparse Attention + Temporal Decay Encoding) + Cross-Modal Attention Extract cross-scale temporal features and financial semantics, dynamically weight and fuse them to improve the predictive power of state representation. Hierarchical adaptive decision Meta-policy network (market state classification) + Multi-execution policy network (scenario-specific policies) This enables human-like reasoning that "first identify the scene, then decide on the action," with strategies that can be explicitly switched. Risk Embedded Joint Optimization Differentiable constrained network (linear inequality projection layer) + risk-adjusted reward function Actions 100% meet risk control requirements; end-to-end strategy training achieves joint optimization of benefits and risks. Online learning of population evolution Parallel exploration across multiple replicas + meta-controller selection (20% culling, top 20% of replicated mutations) + policy gradient updates Avoid getting stuck in local optima with a single strategy, accelerate convergence, and adapt to market pattern drift. Beneficial effects

[0029] Compared with the prior art, the present invention has the following beneficial effects: More comprehensive signals and higher prediction accuracy: By fusing sentiment and event signals from unstructured text and combining improved TCN and Transformer, the predictive power of market state representation is enhanced. Experiments show that, for CSI 300 constituent stocks, the accuracy of predicting the next day's price movement is improved by approximately 12 percentage points compared to using only volume and price data.

[0030] The strategy exhibits strong adaptability and high robustness: the hierarchical reinforcement learning framework enables the system to automatically switch sub-strategies when market conditions change abruptly. In multiple extreme market tests, the maximum drawdown of the strategy is reduced by an average of more than 40% compared to traditional reinforcement learning strategies.

[0031] Intelligent risk management with zero violations: An embedded differentiable constraint network ensures that every output action meets preset risk control limits (industry concentration, turnover rate, value at risk, etc.). Zero-violation instruction output is achieved in live trading simulations, avoiding transaction delays and wasted computing resources caused by traditional post-filtering.

[0032] High degree of automation and shortened development cycle: The entire process from data input to strategy evolution is automated, reducing the strategy iteration cycle from several days of manual parameter tuning to hours. The population evolution mechanism enables the strategy to continuously adapt to market changes without human intervention.

[0033] The risk-reward ratio is significantly improved by adopting the risk-adjusted reward function Rt=rt−λ⋅MDDt−γ⋅TCt. Rt = rt - λ ⋅MDD t - γ ⋅TC t Optimization was performed, and the Sharpe ratio was improved from 0.8~1.2 in the traditional method to 1.5~2.0 in the backtesting, and the Kalmar ratio was improved by more than 2 times. Attached Figure Description To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Figure 1 System architecture diagram. Figure 2. Workflow diagram of the hierarchical reinforcement learning strategy generator. Figure 3 Diagram of risk constraint closed loop.

Claims

1. A quantitative investment strategy generation system, characterized in that, include: A multimodal financial data acquisition and fusion module includes a data acquisition unit and a data fusion unit. The data acquisition unit is configured with several data access adapters to acquire structured data (such as market data and fundamental data) and unstructured data (such as news and reports). The data fusion unit synchronizes the structured and unstructured data through a timestamp alignment device and outputs a unified preprocessed data stream. The deep feature extraction module includes a temporal convolutional network (TCN) and a Transformer encoder. The TCN processes structured time-series data, and the Transformer encoder processes text data. The outputs of both are weighted and fused by an attention fusion unit to form a market state vector. The hierarchical reinforcement learning strategy generation engine includes a meta-policy network and several execution policy networks. The output of the meta-policy network is coupled to the activation input of the execution policy network via a connection line. The output of the execution policy network outputs a trading action signal via a signal scheduling unit. The risk-constrained trading execution module includes a constraint network. The constraint network internally has a weight matrix and threshold parameters. The input of the constraint network is connected to the trading action signal output. The constraint network performs matrix operations on the input signal according to preset risk management rules, compares the threshold, and outputs a compliance check result. The online adaptive learning module includes a meta-controller, an evaluation unit, and a parameter synchronization unit. The meta-controller periodically obtains policy performance data from the hierarchical reinforcement learning policy generation engine. The evaluation unit instantiates several policy copies in parallel in a simulated environment and calculates the Sharpe ratio or Kamma ratio. The parameter synchronization unit writes the preferred parameters into the parameter storage units of the meta-policy network and the execution policy network through a configuration interface based on the evaluation results. The control unit includes a timing scheduler, a data router, and a parameter synchronizer. The output of the timing scheduler is connected to the clock input of each module to achieve timing coordination. The data router forwards the output data stream of each module to the input of the target module through the bus interface. The parameter synchronizer realizes real-time updating of the parameters of each module through shared memory.

2. A method for generating a quantitative investment strategy applied to the system described in claim 1, characterized in that, The process includes the following steps: using the acquisition unit of the multimodal financial data acquisition and fusion module as described in claim 1 to acquire structured time-series data and unstructured text data, and then using the fusion unit of the module to synchronize and output a unified preprocessed data stream through a timestamp alignment device; The structured temporal data is convolved using the TCN in the deep feature extraction module of claim 1, and the text data is self-attention encoded using a Transformer encoder. Then, a market state vector is generated by weighted fusion through an attention fusion unit. This market state vector is input to the meta-policy network input of the hierarchical reinforcement learning policy generation engine of claim 1. The output of the meta-policy network drives the activation input of the corresponding execution policy network. The execution policy network generates a trading action signal via a signal scheduling unit. This trading action signal is input to the constraint network input of the risk-constrained trading execution module of claim 1. The constraint network performs matrix operations based on preset risk management rules and compares thresholds, outputting a compliance judgment result. The compliance-judged trading action signal is executed by the execution unit. Environmental feedback is collected by the monitoring unit and sent to the parameter update module. The parameter update module updates the weights of the hierarchical reinforcement learning model using gradients based on the collected return and risk indicators.

3. The quantitative investment strategy generation system according to claim 1, characterized in that, The deep feature extraction module includes a temporal convolutional network (TCN) and a Transformer encoder.

4. The system according to claim 3, wherein the deep feature extraction module employs an attention mechanism to perform weighted fusion of features from different modalities.

5. The system according to claim 1, characterized in that, The hierarchical reinforcement learning policy generation engine uses a risk-adjusted reward function as its reward function.

6. The system according to claim 5, wherein the reward function formula is: Reward = Stage return rate - λ·Maximum drawdown - γ·Transaction cost.

7. The system according to claim 1, characterized in that, The online adaptive learning module uses a meta-controller to periodically evaluate the Sharpe ratio or Kalmar ratio of each policy replica and perform parameter synchronization for survival of the fittest.

8. The system according to claim 1, characterized in that, When the risk-constrained transaction execution module fails the verification, it includes a projection unit. The projection unit implements a gradient descent algorithm to project non-compliant transaction action signals to transaction actions that meet the constraints based on preset feasible region constraints.

Images