Multi-Modal Deep Learning Architecture for Stock Price Prediction

Stock prediction is one important research problem of applying information technology in financial applications. In general, the conventional approach is usually based on some fundamental and technical analysis methods to analyze and forecast stock process. Recently, text mining has been shown its potential in improving the performance of stock prediction models. In particular, it is based on performing the natural language processing step to extract related textual features from web-based financial news articles for constructing the prediction models. On the other hand, deep learning techniques, which were originally proposed to solve computer vision problems, have been used for text and time series data domain problems. For stock prediction, some related studies have considered deep learning techniques to learn financial news articles and stock charts. However, financial news collected from different web sources or platforms can lead to different textual feature representations, so that the prediction models can perform differently. Therefore, the aim of the first year research is to examine the effect of using financial news collected from different online sources on the performances of different prediction models based on supervised learning classifiers and deep learning networks. For the second year research, related low-level image features are extracted from the stock charts where supervised learning classifiers and deep learning networks are trained and tested for performance comparison. Finally, the aim of the third year research is to employ the ensemble learning technique to construct multi-modal prediction models based on technical indicators, textual features, and image features in order to find out the best combination of multiple prediction models.

Project ID:PB10907-3525
External Project ID:MOST109-2410-H182-012