Convolutional Neural Network for Stock Price Prediction Using Transfer Learning

Abstract

The goal of this paper is to build a trading algorithm by applying image recognition neural network - Convolutional Neural Network(CNN) - to the 2D technical candle stick charts. First, this paper shows a research survey of the previous paper. Second, this paper explains the basic theory of CNN model and how it can works on chart images. Next, this project performs an experimental study of CNN on S&P 500 index from January 1, 1985 to June 30, 2020. The CNN model structure used in this paper is transferred from inception v3 with three additional layers, and the technical indicators used in the input chart image are simple moving average (25 days). The label data used in the model are categorical - either up, flat, or down. The model has 50% accuracy on the test set when conducting three-days ahead forecast, which is higher than the simple momentum strategy and contrarian strategy, indicating its high alpha generating potential. One-day ahead forecast and five-days ahead forecast have lower accuracy than the three-days forecast. This means you might have the best performance when you close your position at T + 3.

Keywords: trading algorithm, convolutional neural network, candle sticks, transfer learning

JEL Classification: C45