Sentiment hedging: Trending stocks on reddit

Abstract

The present document studies the hedging of portfolios in the face of social media sentiment. To do this, a portfolio of stocks is constructed whose returns are correlated with innovations in social media sentiment. To estimate sentiment, two approaches are presented: the first with supervised analysis using the Multinomial Inverse Regression (MNIR) model (Taddy, 2013a) and the second with the lexicon-based Valence Aware Dictionary for Sentiment Reasoning (VADER) model (Hutto and Gilbert, 2014). To estimate sentiment, a database of 56.5 million comments and posts from Reddit is used. To generate the hedging asset allocation, Deep Reinforcement Learning is implemented, specifically the Adaptive Deep Deterministic Policy Gradient (Adaptive DDPG) algorithm (Li et al., 2019). It is found that the way sentiment is estimated determines the algorithm¿s hedging performance to a large extent. For the portfolio that is covered against sentiment estimated by MNIR, notable out-of-sample results are achieved. To highlight the versatility of the methodology, algorithm performance is presented by focusing on maximizing returns. In summary, the best performing portfolios, in terms of return, were those that incorporated social media sentiment analysis. Additionally, portfolios utilizing the Adaptive DDPG algorithm showed a better Sharpe ratio compared to an Equal Weighted Portfolio and individual stock investments.