Technical Summary: FinMamba

Problem Definition

Stock movement prediction is a critical yet challenging task in quantitative trading due to the high volatility of financial data, low signal-to-noise ratios, and the complex, multifaceted nature of market influences. Existing deep learning approaches face two primary limitations:

1. Static or Biased Inter-Stock Relationships: Current models often rely on static prior knowledge (e.g., industry sectors) or purely data-driven posterior correlations (e.g., Pearson coefficients) without adequately accounting for dynamic market conditions. Specifically, they fail to capture how inter-stock correlations intensify during market downturns (systemic risk) and weaken during upturns, a phenomenon driven by macroeconomic factors.
2. Inefficiency and Pattern Recognition: Traditional sequence models like Transformers struggle with long historical data due to quadratic computational complexity and often overemphasize outliers rather than recurring patterns. Conversely, they lack the ability to efficiently model dependencies across multiple time scales (from micro-level minute fluctuations to macro-level economic trends) while maintaining the low latency required for real-time trading.

Methodology

The authors propose FinMamba, a framework integrating a Market-Aware Graph (MAG) and a Multi-Level Mamba (MLM) architecture to address these limitations.

1. Market-Aware Graph (MAG)

The MAG module dynamically models inter-stock relationships by fusing static prior knowledge with dynamic posterior correlations, refined by market index feedback.

• Dynamic Stock Correlation Graph Generation: The model constructs an adjacency matrix $A_t$ by combining:
  • Posterior Correlations ($Q$): Calculated daily using the Spearman coefficient over a lookback window $L$ to capture short-term sequence similarities.
  • Prior Relationships ($D$): An "Industry Decay Matrix" that encodes long-term structural relationships based on primary and secondary industry classifications, assigning higher weights to stocks within the same industry.
  • The final adjacency matrix is computed as $A_t = Q_t \cdot D$.
• Market-Aware Graph Sparsification: To adapt to market volatility, the model employs a pruning mechanism. It calculates a sparsity level $\kappa_t$ based on the market index $M_t$ (represented by the mean of the target stock set). Using an Inception block and a Sigmoid function, the model determines how many edges to retain.
  • Mechanism: When the market index is low (high volatility/crisis), the model retains more edges (higher correlation). When the index is high, it prunes more edges, reflecting the divergence of stock performance.
• Graph Attention Aggregation: A multi-head attention mechanism aggregates messages from neighboring nodes in the pruned graph to generate refined stock embeddings.

2. Multi-Level Mamba (MLM)

The MLM module processes the stock embeddings to capture temporal dependencies across different granularities.

• Multi-Level Projection: The input is projected onto $k$ distinct levels via linear mappings. This allows the model to simultaneously extract fine-grained local details (micro-level) and broader global patterns (macro-level).
• Selective State Space Mechanism: Each level is processed by a Mamba block. Unlike Transformers, Mamba utilizes a selective mechanism that allows the model to recall specific historical patterns relevant to the current input while discarding irrelevant information. This enables the model to handle long sequences with linear time complexity and memory consumption.
• Output: The outputs from all levels are re-projected, concatenated, and passed through a linear layer to predict the ranking scores for the next trading day.

3. Optimization

The model is trained end-to-end using a combined loss function:

• Rank & Prediction Loss ($L_{RP}$): A combination of Mean Squared Error (MSE) for point-wise regression and a pairwise ranking loss to optimize the relative ordering of stocks.
• Graph Information Bottleneck (GIB) Loss ($L_{GIB}$): A regularization term designed to minimize the mutual information between the aggregated graph embeddings and the original input, ensuring the model retains key correlation information while filtering out noise.

Key Contributions

The paper claims the following contributions:

• Market-Aware Graph Construction: This is the first work to explicitly enhance inter-stock relationship modeling under specific macroeconomic market conditions. It integrates static priors (industry) and dynamic posteriors (price sequences) and refines them using market index feedback to adaptively adjust correlations based on market trends.
• FinMamba Framework: The proposal of a novel architecture that combines the Market-Aware Graph with a Multi-Level Mamba. This framework effectively models pattern-oriented dependencies across multiple time scales (micro to macro) while maintaining high computational efficiency.
• State-of-the-Art Performance: Extensive experiments on U.S. (S&P 500, NASDAQ 100) and Chinese (CSI 300, CSI 500) stock markets demonstrate that FinMamba achieves superior prediction accuracy and trading profitability compared to existing quantitative investment methods and general time-series forecasting models, all while maintaining low computational complexity.

Experimental Results

The authors evaluated FinMamba against a wide range of baselines, including classic quantitative strategies (BLSW, CSM), deep reinforcement learning methods (AlphaStock), and various deep learning architectures (Transformer, Mamba, THGNN, PatchTST, etc.).

• Performance Metrics: The model was assessed using Annual Return Ratio (ARR), Annual Volatility (AVol), Maximum Draw Down (MDD), Annual Sharpe Ratio (ASR), and Information Ratio (IR).
• Findings:
  • FinMamba achieved the best or second-best results across all four datasets on most metrics, particularly in ARR, ASR, and IR.
  • It outperformed general time-series forecasting models (e.g., Crossformer, iTransformer) which lack graph-based inter-stock relationship modeling.
  • It surpassed other graph-based methods (e.g., THGNN, VGNN) by effectively incorporating market index feedback into the graph structure.
  • Ablation Studies: Removing the Industry Decay Matrix, Market-Aware Sparsification, or the Multi-Level Mamba structure resulted in significant performance degradation, confirming the necessity of each component.
  • Efficiency: FinMamba demonstrated lower inference time and memory usage compared to Transformer-based models, especially as the lookback horizon increased, validating its linear complexity advantage.
  • Case Study: Analysis of the 2023 UAW strike showed that FinMamba successfully captured the evolving correlation dynamics between traditional automakers (GM, Ford) and Tesla, reflecting the shift from distinct market positions to intensified competition.

Significance and Claims

The paper positions FinMamba as a significant advancement in quantitative trading by addressing the dual challenges of dynamic market adaptability and computational efficiency. The authors claim that by integrating macroeconomic market feedback directly into the graph construction process, the model can better understand market dynamics, particularly during periods of volatility. Furthermore, the use of the Multi-Level Mamba architecture allows for the efficient extraction of multi-scale temporal patterns without the computational burden of attention mechanisms, making the approach suitable for real-time algorithmic trading. The work suggests that combining graph-based structural priors with the selective state space modeling of Mamba offers a robust paradigm for financial time series prediction.