Abstract
Stock market forecasting is a challenging task due to the low signal-to-noise ratio and the non-stationary nature of financial data. This paper presents the Automatic De-Biased Temporal-Relational Model (ADB-TRM), a framework designed to enhance stock recommendation robustness. ADB-TRM utilizes a bilevel optimization meta-learning process integrated with temporal and relational adversarial training to mitigate inherent biases and distribution shifts. Evaluations on major global indices demonstrate that ADB-TRM significantly outperforms existing state-of-the-art models, achieving up to a 28% improvement in cumulative returns.
1. Introduction
Predicting stock market movements is notoriously difficult because the underlying "signals" are often obscured by significant "noise." Traditional Temporal-Relational Models (TRMs) analyze stock trends and inter-stock relationships but frequently succumb to three primary challenges: temporal bias, relational bias, and distribution shifts.
As illustrated in Figure 1, the statistical distribution of market data changes over time (e.g., comparing 2022 to 2023). This non-stationarity breaks the independent and identically distributed (i.i.d.) assumption, causing models trained on historical data to fail in new market regimes.
2. Methodology
The ADB-TRM framework employs a bilevel optimization approach to decouple local de-biasing from global pattern recognition.
2.1 Inner Loop: De-biasing and Adversarial Training
The inner loop (Figure 3) focuses on identifying dynamic relationships between stocks using Adaptive Graph Learning. To ensure robustness against noise, the model incorporates two adversarial strategies:
- Temporal Adversarial Training (TAT): By injecting noise from various time periods, the model learns to distinguish stable signals from transient fluctuations.
- Relational Adversarial Training (RAT): This component generates adversarial graph structures to prevent the model from over-relying on "hub" stocks, thereby mitigating relational bias.
2.2 Outer Loop: Meta-Learning for Invariance
The outer loop utilizes a WaveNet-based architecture to analyze global market sequences. It aims to extract "relative invariant" embeddings—features that remain consistent despite shifting market distributions—allowing the model to adapt to regime changes effectively.
3. Experimental Results
The model was evaluated on the NASDAQ, NYSE, and TSE datasets. ADB-TRM consistently outperformed 16 baseline methods, including traditional regression and GNN-based models.
Key performance metrics include:
- Profitability: An average relative improvement of 28.41% in cumulative returns (IRR).
- Risk-Adjusted Returns: A 9.53% improvement in Sharpe Ratio across datasets. On NASDAQ, the model achieved a Sharpe Ratio of 1.66.
- Generalization: The model showed strong performance across different market sizes and conditions, from the large-scale NYSE to the more concentrated TSE.
4. Conclusion
ADB-TRM addresses the critical issues of noise and bias in financial time-series modeling. By combining adversarial training with a meta-learning framework, it identifies robust signals that persist across market cycles. While the training process is computationally intensive, the model's efficiency during inference makes it a practical tool for real-time stock recommendation. Future research will explore the integration of alternative data sources to further refine the relational graph.
Summary updated