ICAIMT: ART-LSTM: Augmented Reverse Training for Data-Efficient Time Series Forecasting

Financial time series forecasting faces significant challenges due to data scarcity, high volatility, and inherent nonlinearities. Complex deep learning models, such as transformers, typically require extensive datasets and computational resources, making them prone to overfitting in financial contexts where datasets are limited. To address this, we propose Augmented Reverse Training LSTM (ART-LSTM), a novel data augmentation strategy for time series forecasting using a straightforward unidirectional LSTM architecture. ART-LSTM leverages both forward and reversed sequences during training, effectively doubling the available training data without increasing architectural complexity. Our approach maintains computational simplicity while enhancing model robustness and generalisation. Empirical evaluations on challenging datasets, including daily S&P 500 index prices and USD/EUR exchange rates, demonstrate that ART-LSTM consistently outperforms traditional statistical methods (ARIMA), standard recurrent neural networks (RNN, GRU, and LSTM), and multi-layer perceptrons (MLPs), achieving substantial reductions in Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE). Overall, ART-LSTM provides a practical and data-efficient solution for financial forecasting tasks characterised by limited data availability and volatile dynamics.