Date of Award
8-2021
Document Type
Open Access Dissertation
Degree Name
Doctor of Philosophy (PhD)
Department
Business Administration
First Advisor
Josephine M. Namayanja
Second Advisor
Sangwan Kim
Third Advisor
H. Zafer Yuksel
Abstract
Recent studies in asset pricing find that Artificial Neural Networks (also known as Deep Learning models) provide the most accurate firm-level return predictions using a vast set of predictive signals. These models offer high predictive accuracy over long out-of-sample periods, translating into highly profitable trading strategies. In this thesis, I argue that sentiment-driven mispricing is a vital source of the high predictability and the resulting profitability implied by deep learning models. Using a novel Artificial Neural Network (ANN) regression model, I obtain firm-level predictions conditional on 54 firm-level characteristics and on an encoded representation of the macro-economic state. These predictions provide important insights into the sources of overall cross-sectional return predictability. First, the future negative returns are predictable out-of-sample which implies negative expected returns. Such predictability is hard to reconcile with a risk-based explanation. Secondly, the predictability in negative returns is higher following periods of high sentiment and vice versa. This evidence is consistent with the existence of a market-level investor sentiment that drives misvaluations. Third, a long-short strategy based on ANN prediction deciles is more profitable following periods of high sentiment. This disparity in profitability points to arbitrage asymmetry implied by short-sale constraints. Fourth, the predictability in losses and high profitability of the ANN top decile vanishes in estimation horizons longer than a month. This suggests that mispricing is short-lived and that predictability is realized due to corrections to such misvaluations. These corrections are preceded by high put-to-call(PCR) trading volumes and high implied volatility(VIX). Finally, the short-term and long-term predictions load on different conditioning variables indicating varying sources of predictability across return horizons. Overall, these findings are consistent with the existence of sentiment-driven short-lived mispricing that corrects in longer horizons.
Recommended Citation
Khan, Usama A., "Return Predictability and Market Sentiment: Evidence from Deep Learning" (2021). Graduate Doctoral Dissertations. 699.
https://scholarworks.umb.edu/doctoral_dissertations/699