Stacked Denoising Autoencoder Based Stock Market Trend Prediction via K-Nearest Neighbour Data Selection

In financial applications, stock-market trend prediction has long been a popular subject. In this research, we develop a new predictive model to improve the accuracy by enhancing the denoising process which includes a training set selection based on four K-nearest neighbour (KNN) classifiers to generate a more representative training set and a denoising autoencoder-based deep architecture as kernel predictor. Considering the good agreement between closing price trends and daily extreme price movements, we forecast extreme price movements as an indirect channel for realising accurate price-trend prediction. The experimental results demonstrate the effectiveness of the proposed method in terms of its accuracy compared with traditional machine-learning models in four principal Chinese stock indexes and nine leading individual stocks from nine different major industry sectors.