Trading by Charts: A Multivariate CNN System to Predict Retail Investor Trading

Understanding retail investor behavior is interesting and important but difficult, the prediction of retail
investor trading is even much harder. This dissertation sheds some light on predicting retail investors’
holding on stocks by building an accessible time-series prediction system using the CNN technique
for pre-COVID and COVID-19 periods. For the first time, a multivariate convolutional neural network
(CNN) is built with numerical and graphical data. Focusing on the U.S. stock markets between 2018
and 2020, this study utilizes component stocks of the S&P 500 index as the sample with relative data
on stock characteristics, retail investor holding, and retail investor ownership. The pioneering
multivariate CNN system performs great in predicting retail investor trading and outperforms the
random forests models built which only apply numerical data. The results support previous studies on
the performances of deep learning techniques like CNN and investor trading behavior and sentiment.
Besides, retail investor holding contains little predictive information for stock price movement. This dissertation contributes to the economic and financial literature by filling the gap in the predictions of retail investor behavior using cutting-edge machine learning techniques based on novel applications of data. In addition, this prediction system can improve social welfare by helping retail investors make less biased decisions, informing financial institutions to better engage with retail investors, and assisting financial authorities to better monitor and manage risks caused by retail investors in the market.