AComNN: Attention enhanced Compound Neural Network for financial time-series forecasting with cross-regional features

Abstract

In recent years, many works spring out to adopt the forecast-based approach to support the investment decision in the financial market. Nevertheless, most of them do not consider mining the hidden patterns in the cross-regional financial time-series. However, the fluctuation in financial markets has always been affected by the global economy, instead of a single market. To overcome this issue, this article proposes an Attention enhanced Compound Neural Network (AComNN) that can be applied on features of multiple-sources, including different financial markets and economic entities. The proposed novel approach compounds of Artificial Neural Network (ANN), Long Short-Term Memory (LSTM), and self-attention to progressively capture the time-zone-dependent context behind the financial time-series across regions with multiple filters. Thereby, it provides trading signals for supporting the financial investment decision. The proposed AComNN has been applied on the Hong Kong Hang Seng Index (HSI) trend prediction based on various initial features across regions. The experimental result demonstrates that the AComNN achieves the highest average accuracy for the one-day ahead trend prediction over 60%. Besides, it reveals highly superior competitiveness on the forecasting capability improved by 13.36% on average compared with the baselines. Therefore, we encourage to adopt the proposed method to the practitioners and provide a new thought, considering the analysis of cross-regional features, in the financial time-series forecasting.