Context-aware Embeddings for Stock Prediction with Visual Clues

The stock market is a complex dynamic system, and its fluctuation must be the result of the joint influence of various market factors. Most previous studies usually attempt to predict stock movements more accurately by involving more market information regarded as a forecast indicator vector, which inevitably ignores feature interactions and causes significant gaps between the real-world deployment and such accuracy breakthrough. In this study, we propose a dynamic fusion framework to learn the interactions according to the prediction tasks, and innovatively interpret predictive models with visual clues from both the input perspective and the output perspective. Specifically, a core-tensor parameter is utilized to bridge the representations learning and model predicting, which is able to thoroughly fuse different factors and preserve their interactions. An interpretation design is further adopted to identify the key factors, the synergistic effects, and the profitable outputs. In addition, for the unstructured input, a context-aware hierarchy-attention interpretation is proposed to gain insight into the multi-granularity textual contents. Experiments performed on the real datasets demonstrate the superiority of the proposed approach over several strong baselines. A case study is further conducted to explain what can be interpreted out of the proposed fusion framework.