Improving exchange rate forecasting via a new deep multimodal fusion model.

Exchange rates are affected by the impact of disparate types of new information as well as the couplings between these modalities. Previous work mainly predicted exchange rates solely based on market indicators and therefore achieved unsatisfactory results. In response to such an issue, this study develops an inventive multimodal fusion-based long short-term memory (MF-LSTM) model to forecast the USD/CNY exchange rate. Our model consists of two parallel LSTM modules that extract abstract features from each modality of information and a shared representation layer that fuses these features. In terms of the text modality, bidirectional encoder representations from transformers (BERT) is applied to conduct a sentiment analysis on social media microblogs. Compared to previous studies, we incorporate not only market indicators but also investor sentiments into consideration, treating the two types of data differently to match their exclusive characteristics. In addition, we apply the multimodal fusion technique and contrive a deep coupled model rather than a shallow and simple model to reflect the couplings between the two modalities. As a consequence, the experimental results obtained over a 15-month period exhibit the superiority of the proposed approach over nine baseline algorithms. The purpose of our study is to demonstrate that it is practicable and effective to incorporate multimodal fusion into financial time series forecasting.