An Adaptive Multimodal Learning Model for Financial Market Price Prediction

Investors' trading behavior is influenced by a multimode of information sources such as technical analysis, news dissemination, and sentiment, which results in the non-stationary behavior of financial time series. With advancements in deep learning, studies considering temporal relationships in each data mode and applying heterogeneous data fusion techniques for market prediction are increasing. While net price change prediction is helpful for investors, most previous deep learning models only predict the up/down trend of price as the non-stationary behavior of price time series influences the regression performance. In this work, we present an adaptive model for price regression, which learns interdependencies between the distribution of multimode data and the amount of price change around an average price for snapshots of systems. We use news content, the mood in specialized newsgroups, and technical indicators for data representation. Different news topics, also known as modalities, can be absorbed by investors with different diffusion speeds; hence we use a concept-based news representation method that reflects news topics in a news vector. Also, our model considers the positive/negative mood in specialized newsgroups and technical indicators. To capture complex temporal characteristics in the distribution of economic concepts in the news sequence, we use a recurrent convolutional neural network and other recurrent layers to perceive changes in technical indicators and mood in specialized newsgroups. In the fusion layer, our model learns to normalize data points based on their estimated distribution and the importance weight of each data mode to handle multimodality challenges. To overcome the non-stationary behavior of price, we let the network learn how to drift the predicted values around the average price of that packet. Our experiments demonstrate a significant 40.11% error reduction compared to the baselines. We also discuss the adaptability, and price prediction capability of our proposed approach.