A New Neural Network Approach for Predicting the Volatility of Stock Market

The prediction of stock market volatility is an important and challenging task in the financial market. Recently, neural network approaches have been applied to obtain better prediction of volatility, however, there have been few studies on artificial manipulation of the volatility distribution. Because the probability density of volatility is extremely biased to the left, it is a challenging problem to obtain successful predictions on the right side of the density domain, that is, abnormal events. To overcome the problem, we propose a novel approach, we call it Volume-Up (VU) strategy, that manipulates the original volatility distributions of invited explanatory variables including the Standard & Poor’s 500 (S&P 500) stock index by taking a non-linear function on them. Multi-Layer Perceptron (MLP), Recurrent Neural Network (RNN), and Long Short-Term Memory (LSTM) are used as our implementation models to test the performances of VU. It is found that the manipulated information improves the prediction performance of one day ahead volatility not only on the left but also on the right probability density region of S&P 500. Averaged gains of root mean square error (RMSE) and RMSE on $$P>0.8$$ P > 0.8 against the native strategy over all the three models were 27.0% and 19.9%, respectively. Additionally, the overlapping area between label and prediction is employed as an error metric to assess the distributional effects by VU, and the result shows that VU contributes to enhance prediction performances by enlarging the area.