DoubleAdapt: A Meta-learning Approach to Incremental Learning for Stock Trend Forecasting

TL;DR

DoubleAdapt tackles incremental stock trend forecasting under distribution shifts by jointly learning data adaptation and model initialization through a bi-level meta-learning framework. The data adapter reshapes features and labels to form a locally stationary distribution, while the model adapter provides a task-specific initialization enabling rapid adaptation to incremental data. Empirical results on CSI 300 and CSI 500 demonstrate state-of-the-art predictive performance and favorable online efficiency, with ablations validating the contribution of both adapters and the multi-head design. The approach offers a practical path to robust, frequent online updates in non-stationary markets, and can be extended to hybrids with rolling retraining for sustained performance.

Abstract

Stock trend forecasting is a fundamental task of quantitative investment where precise predictions of price trends are indispensable. As an online service, stock data continuously arrive over time. It is practical and efficient to incrementally update the forecast model with the latest data which may reveal some new patterns recurring in the future stock market. However, incremental learning for stock trend forecasting still remains under-explored due to the challenge of distribution shifts (a.k.a. concept drifts). With the stock market dynamically evolving, the distribution of future data can slightly or significantly differ from incremental data, hindering the effectiveness of incremental updates. To address this challenge, we propose DoubleAdapt, an end-to-end framework with two adapters, which can effectively adapt the data and the model to mitigate the effects of distribution shifts. Our key insight is to automatically learn how to adapt stock data into a locally stationary distribution in favor of profitable updates. Complemented by data adaptation, we can confidently adapt the model parameters under mitigated distribution shifts. We cast each incremental learning task as a meta-learning task and automatically optimize the adapters for desirable data adaptation and parameter initialization. Experiments on real-world stock datasets demonstrate that DoubleAdapt achieves state-of-the-art predictive performance and shows considerable efficiency.

Figures (7)

• Figure 1: Illustration of distribution shifts in CSI 300 stock set. The vector of each stock including dozens of technical indicators and the corresponding label is mapped via t-SNE t-SNE to a 1-D point on the horizontal axis. We estimate the distributions with a kernel density estimator. We plot the distribution of the incremental data in one month in purple, the distribution of all the previous data in black, and the distribution of the next month's data in green.
• Figure 2: Illustration of IL for stock trend forecasting. $r$ is the timespan of incremental data or test data.
• Figure 3: Illustration of data adaptation. The function $G$ adapts the feature distributions of the incremental data and the test data to an agent feature distribution. Similarly, the function $H$ adapts the two posterior distributions to an agent one. Its inverse function $H^{-1}$ restores the model outputs to the original posterior distribution of the test data.
• Figure 4: Overview of DoubleAdapt with a data adapter ${DA}$ and a model adapter ${MA}$. The parameters are shown in red.
• Figure 5: Performance comparison under different hyperparameters on CSI 300. Outliers are marked in diamonds.