Few-Shot Financial Time Series Forecasting via Meta-Learning and Temporal Attention

Abstract

In real-world applications, there are challenges to forecasting a financial time series due to its natural non-stationarity, presence of noise, and lack of labeled data. In particular, it is often found that newly listed or low-liquidity assets have significant data shortage problems, making it difficult to effectively generalize the commonly used deep learning model. This paper address this issue by casting the problem of financial forecasting as a few-shot learning setting and propose a meta-learning approach to improving the adaptability of the models under few-shot settings. Specifically, a framework based on a Transformer is adopted as the main model to learn the temporal long-range dependency in multivariate time series. A meta-learning approach is used to learn a transferable parameter initialization on different financial tasks, to improve its efficiency in low-data scenarios. This enables the model to adapt to new assets very fast, based on a very limited set of support samples. The proposed approach is tested on the Huge Stock Market Dataset where each stock is a task in different few shot setting (K=1, 3, 5, 10). Experimental results demonstrate that the proposed Meta-Baseline significantly outperforms the conventional baselines (including Transformer, LSTM and GRU) in each case. In the toughest scenario (K=1), the suggested approach decreases the prediction error by more than 70% when compared to typical Transformer models. Furthermore, more support samples in the model result in steady and consistent improvement in stability of the model. These findings highlight the effectiveness of meta-learning for few-shot financial time series prediction and suggest its potential for real-world applications in data scarce environments.