Predicting stock prices or their directional movement has long been regarded as one of the most challenging problems of modern time-series prediction. Understanding China’s underrepresented A-share market is of growing interest to the global equity market with its rising investment inflows into one of the largest equity markets in the world, which thus becomes our intuitive motivation. We have several objectives in this study: 1) We aim to discover whether time is a factor under high-frequency data at one-minute intervals for the banking sector in China’s A-share market using the classic time-series model. 2) We aim to discover the importance of the technical analysis indicators using high-frequency data at one- minute intervals for the banking sector in China’s A-share market with the classic nonlinear model. 3) We aim to reduce the forecast error rate and to improve the prediction accuracy rate of the stock price forecast and stock price directional movement prediction. To achieve this goal, we adopt the deep bidirectional long short-term memory (BiLSTM) neural network model, a type of deep recurrent neural network (RNN) with “learn” and “forget” capabilities, as well as “learning forward” and “learning backward” capabilities, together with the various deep learning hyperparameter optimization algorithms. 4) Finally, we aim to compare the accuracies of the predictions of stock prices and their directional movement predictions between classic time-series and regression models and the deep BiLSTM model with high- frequency data in China’s A-share market.

We collected the one-minute high-frequency stock data from February 1, 2018 to May 31, 2018; the market had a downward trend during this period due to the beginning of the US- China trade war. In our first research problem, the data suggest that the closing prices of the top Chinese banks in China’s A-share market becomes stationary after differencing once using high-frequency data for all the studied Chinese banks. More than 50% of the studied Chinese banks are found to have time lags of a high-frequency at one-minute intervals stock closing prices, which are regressed on their own for up to two minutes. More than 60% of the studied Chinese banks are found to have a regression error or white noise that is linearly related to the error terms of which their values persist continuously at one minute in the past, and for the remaining studied Chinese banks, this relationship persists for up to 5 minutes in the past. An empirical study shows that deep BiLSTM model achieves an average error reduction rate of 92.3% compared to the autoregressive integrated moving average (ARIMA) model in terms of predicting the high-frequency stock price for the top Chinese banks in China A-share market. In our second research problem, the accuracy rate improvement of the BiLSTM model over nonlinear logistic regression (LOG) analysis is shown to be as high as 21% when predicting stock price directional movement using high-frequency data for the top banks in China’s A- share market. Some technical analysis indicators, including relative strength index (RSI), Hilbert transform oscillator (HTTRENDLINE) and volume traded in shares, are found important drivers of the directional movement of Chinese banks’ high-frequency stock prices. In particular, a positive and highly significant (1% level of statistical significance) link is found between stock price directional movement and the RSI for all Chinese banks with high- frequency data. Our findings enable both practitioners and researchers of China’s A-share market to discover and understand the impacts of different technical analysis indicators on the high-frequency directional movement of the stock prices in the banking sector in China’s A-share market.