Transfer Learning Based Long Short-Term Memory Network for Financial Time Series Forecasting

Abstract

Long Short-Term Memory (LSTM) neural network is widely used to deal with various temporal modelling problems, including financial Time Series Forecasting (TSF) task. However, accurate forecasting of financial time series remains a difficult problem due to its implicit complex information and lack of labeled training data. To alleviate the limitation of overfitting caused by insufficient clean data, a new approach using LSTM based on transfer learning is proposed in our study for financial TSF task, termed as ADA-LSTM for short. Concretely, we not only implement a typical Adversarial Domain Adaptation architecture, but also tactfully introduce a smoothed formulation of Dynamic Time Warping (soft-DTW) in adversarial training phase to measure the shape loss during the transfer of sequence knowledge. Compared to many existing methods of selecting potential source domain during transfer learning in TSF, in our study, appropriate source dataset is selected from a novel perspective using temporal causal discovery method via transfer entropy instead of using statistical similarity across different time series. The feasibility and effectiveness of ADA-LSTM are validated by the empirical experiments conducting on different financial datasets.