Application of Simple Convolutional Neural Networks in Equity Price Estimation

The problem of price estimation is well established in the field of finance, especially speaking of future equity prices. There are many classical approaches to estimate the expected price, such as auto-regression, moving averages, combination of both in the form of ARMA and ARIMA models, and many more. But with the development of technology, completely new techniques arise. In the current study we are attempting to use deep learning techniques, by training a simple (small number of layers) Convolutional Neural Network (CNN) on the graphical representation of the prices, in black and white scale. Normally, CNNs are used to classify data and are not very well suited to generated ‘regression type’ results. For that reason, we are basing our short-term (work week) price estimation on a categorical result. We use the daily standard deviation as a measure in order to split the future price in seven categories. The results show that CNN is able to effectively improve naive estimation of the price. Also, we uncover few problems and possible solutions that arise during the training of the model, related to over-fitting.