Application of BiLSTM-Transformer in Portfolio Optimization

Abstract

Portfolio optimization models that use predictions can effectively capture short-term investment opportunities. However, in traditional models, inaccurate predictions of the expected excess return of different assets can negatively impact investment performance. Deep learning models have demonstrated significant advantages over time series models in this regard. This paper connects Transformer model and the BiLSTM model, which is short for bi-directional Long Short-Term Memory, for return prediction for portfolio model performance enhancement. To be specific, the model of BiLSTM-Transformer is firstly applied for predicting the yield of alternative assets, which is then incorporated in the meanvariance (MV) model. Using 6 component stocks of the US30 index as alternative assets, 270 investments are conducted, and the empirical results are compared with LSTM and Transformer model. The comparison verifies the superiority of BiLSTM-Transformer model in improving prediction accuracy and boost of portfolio model performance.