Incremental Learning of Stock Trends via Meta-Learning with Dynamic Adaptation

TL;DR

The paper addresses stock trend forecasting under concept drift by proposing MetaDA, a meta-learning framework with dynamic adaptation that simultaneously leverages emerging patterns in the latest data and recurring patterns in historical data. It introduces a task inference module and task embeddings to selectively incorporate informative historical data during model adaptation, while updating the forecaster with a single gradient step and performing online updates for subsequent tasks. Empirical results on CSI-300 and CSI-500 demonstrate state-of-the-art performance across multiple forecasters and metrics, with ablation studies validating the contributions of dynamic data selection and the task inference mechanism. The approach offers practical benefits for real-time stock prediction by balancing accuracy and adaptation efficiency, and future work may integrate additional side information such as news sources.

Abstract

Forecasting the trend of stock prices is an enduring topic at the intersection of finance and computer science. Periodical updates to forecasters have proven effective in handling concept drifts arising from non-stationary markets. However, the existing methods neglect either emerging patterns in recent data or recurring patterns in historical data, both of which are empirically advantageous for future forecasting. To address this issue, we propose meta-learning with dynamic adaptation (MetaDA) for the incremental learning of stock trends, which periodically performs dynamic model adaptation utilizing the emerging and recurring patterns simultaneously. We initially organize the stock trend forecasting into meta-learning tasks and train a forecasting model following meta-learning protocols. During model adaptation, MetaDA efficiently adapts the forecasting model with the latest data and a selected portion of historical data, which is dynamically identified by a task inference module. The task inference module first extracts task-level embeddings from the historical tasks, and then identifies the informative data with a task inference network. MetaDA has been evaluated on real-world stock datasets, achieving state-of-the-art performance with satisfactory efficiency.

Figures (4)

• Figure 1: The incremental learning of stock trends, where the forecasting model is periodically updated via model adaptation. In the $i$ th model adaptation, the latest labeled data (training set) $\mathcal{D}_{tr}^i$, the upcoming data (testing set) $\mathcal{D}_{te}^i$, and the historical data $\mathcal{H}^i$ form the $i$ th learning task $\mathcal{T}_i$. We aim to improve the overall performance on testing sets across future tasks. The stock data are also visualized via t-SNE and the results are presented at the bottom of the figure.
• Figure 2: The overview of MetaDA's working pipeline in the $i$ th model adaptation. A set of historical data $\mathcal{D}_{tr}^{h^*}$ is first selected by the task inference module. Model adaptation is then conducted based on $\mathcal{D}_{tr}^{h^*}$ and the latest labeled data $\mathcal{D}_{tr}^{i}$. The forecasting model subsequently predicts stock trend on $\mathcal{D}_{te}^{i}$, and learns the data after the arrival of labels $\mathbf{\Gamma}_{te}^i$.
• Figure 3: The results of Friedman test at 90% significance level, which is a general analysis across different datasets, forecasters, and metrics. If the bars of two methods do not overlap, it indicates that they are significantly different.
• Figure 4: The CSI-300 index (SH000300) of $\mathcal{D}_{tr}^{196}$, $\mathcal{D}_{te}^{196}$, and corresponding $\mathcal{D}_{tr}^{h^*}$, which has a stock price trend similar to that of $\mathcal{D}_{te}^{196}$. With the introduction of $\mathcal{D}_{tr}^{h^*}$, IC has increased by 50 % on this task.