Forecasting Variation Trends of Stocks via Multiscale Feature Fusion and Long Short-Term Memory Learning

Forecasting stock price trends accurately appears a huge challenge because the environment of markets is extremely stochastic and complicated. This persistently motivates us to seek reliable pathways guide trading. While Long Short-Term Memory (LSTM) network has dedicated gate structure quite suitable for prediction based on contextual features, we propose novel LSTM-based model. Also, devise multiscale convolutional feature fusion mechanism model extensively exploit relationships hidden in consecutive time steps. The significance our designed scheme twofold. (1) Benefiting from both long- short-term memories, can use given history data more adaptively than traditional models, which greatly guarantees performance financial series (FTS) scenarios thus profits trends. (2) diversify representation capture FTS essence fairly facilitates generalizability. Empirical studies conducted three classic sets, i.e., S&P 500, DJIA, VIX, demonstrated effectiveness stability superiority suggested method against few state-of-the-art models using multiple validity indices. For example, achieved highest average directional accuracy (around 0.71) employed sets.