Preliminary Study on Interpreting Stock Price Forecasting Based on Tree Regularization of GRU

摘要

The gated recurrent unit (GRU) deep model is interpreted to predict price’s falling or rising. By using a technique called Tree Regularization of Deep Models for Interpretability, a GRU network is converted to a decision tree (called GRU-Tree) to interpret its prediction rules. This approach was tested by experimenting on a few sample stocks (e.g., the Gree company) and a main stock market index (SSE Composite Index) in China. The discovered prediction rules actually reflect a general rule called Mean Reversion in stock market. Results show that the GRU-Tree is more effective (higher AUC) than the decision tree directly trained from the data for small and moderate average path length (APL) of trees. And the fidelity between GRU and its generated GRU-Tree is high (about 0.8).