DoubleAdapt: A Meta-learning Approach to Incremental Learning for Stock Trend Forecasting
DoubleAdapt: A Meta-learning Approach to Incremental Learning for Stock Trend Forecasting
DoubleAdapt addresses the challenge of distribution shifts in incremental learning for stock trend forecasting by introducing two meta-learners: a data adapter and a model adapter. The data adapter transforms features and labels of both incremental and test data to reduce covariate and conditional distribution shifts, making the data more stationary. The model adapter learns a robust parameter initialization for each task, enabling the forecast model to quickly adapt to incremental data while retaining generalization. Each incremental learning task is formulated as a bi-level optimization problem: the lower level fine-tunes the forecast model on adapted incremental data, and the upper level optimizes the adapters using test error. Experiments on real-world Chinese stock market datasets (CSI 300 and CSI 500) demonstrate that DoubleAdapt achieves state-of-the-art predictive performance in terms of information coefficient (IC), RankIC, and excess annualized return, while being computationally efficient compared to rolling retraining. The framework is flexible and can be combined with various neural network architectures for stock trend forecasting.
Highlights
- 1Proposes DoubleAdapt, an end-to-end incremental learning framework with data and model adapters to handle distribution shifts in stock trend forecasting.
- 2Formulates each incremental learning task as a bi-level optimization problem, where lower-level fine-tunes the forecast model and upper-level optimizes meta-learners.
- 3Introduces a data adapter that adapts features and labels to reduce covariate and conditional distribution shifts between incremental and test data.
- 4Introduces a model adapter that learns a robust parameter initialization for each task, balancing adaptiveness and generalization.
- 5Achieves state-of-the-art predictive performance on real-world stock datasets with high efficiency compared to rolling retraining.
Methods
- MMeta-learning with bi-level optimization: lower-level fine-tuning of forecast model on adapted incremental data, upper-level optimization of data and model adapters using test error.
- MData adaptation via feature adaptation layer G and label adaptation layer H (with inverse H^{-1}) to align distributions.
- MModel adaptation via a meta-learner that provides task-specific parameter initialization for the forecast model.
- MOnline incremental learning with periodic tasks, using one-step gradient descent for efficiency.
Results
- RDoubleAdapt outperforms rolling retraining and other meta-learning baselines in terms of predictive accuracy (e.g., IC, RankIC, excess annualized return) on CSI 300 and CSI 500 datasets.
- RThe data adapter effectively mitigates both gradual and abrupt distribution shifts, as shown in t-SNE visualizations.
- RThe model adapter provides a good initialization that balances adaptiveness and robustness, leading to better generalization on test data.
- RDoubleAdapt is significantly more efficient than rolling retraining, with lower time and space complexity.
- RAblation studies confirm the necessity of both data and model adapters for optimal performance.
Analyze Paper
Generate insights from "DoubleAdapt: A Meta-learning Approach to Incremental Learnin...".