Stock Price Forecasting: New Model for Stocks Selection and Price Forecasting Based on Convolutional Neural Network

Abstract

Stock price forecasting is a difficult task given the high volatility of data, at the same time investors always need to anticipate price changes to maximize profit or minimize losses, which justifies the number of works carried out or in progress in this field. In this paper, we present a new model for selecting profitable stocks with low risk and forecasting close prices for a given horizon. Our model is a succession of three phases: first phase is purely dedicated to data cleaning preprocessing and return calculation, then the second phase of selecting profitable stocks with relatively low risk based on Sharpe ratio. The last phase is for training and testing the CNN model we start our training by a small number of epochs and control the error rate for each stock price prediction, stock with an error rate above the error threshold will be discarded and we will increase the epoch number and reduce the threshold error to keep only stocks that the model predict with high accuracy. The model is designed for short term forecasting. The obtained results are very satisfactory and error rate is very reduced.