Classifying stock returns using high-frequency fundamental factors and convolutional neural networks

Sammanfattning: We evaluate the usefulness of high-frequency fundamental factor exposures of five US equities, between 2013 and 2017, as features for classifying and predicting the binary movements of the same stocks in 5-minute and 20-day intervals using Convolutional Neural Networks (CNN). After plotting rolling factor betas (Market, HML, SMB) and the close price of a given stock in the corresponding intervals, these time series are converted into images as Gramian Angular Difference Fields (GADF) and then concatenated to be fed to the CNN as input. Two types of convolutional neural networks are trained on these images and used for a binary classification task of determining whether the close price is likely to increase or decrease in the consecutive time unit. For comparison, the same analysis is conducted with technical indicators (RSI, EMA, %K) and a combination of the two (Market, HML, RSI). The results of this paper show moderate performance of the trained CNN, achieving a maximum accuracy on test data of 54.7% for a 20-day interval using images with a combination of both technical indicators and fundamental factors. For further research, we suggest using a longer forecast and classification horizon, and exploring alternate ways to linear regression for high-frequency beta estimation for fundamental factors.