This thesis consists of three related studies investigating the practical economic
utility of news analytics in momentum-style equity portfolios. We employ a
dataset of 5.3 million news items relating to historical constituents of the S&P
500 stock index from 2003 to 2018 to examine exploitable relationships between news media and stock returns over one - six month horizons.
The first study inspects the predictive capacity of news media at the stock level
through firm-specific regression analyses. Tests include single period regressions subject to a variety of subset and specification robustness tests, extension to multiperiod forecast horizons, and VAR cross-dependency analysis. With rates of statistical significance comparable to pure noise covariates and weak effect sizes, we fail to find support for the hypothesis that either news sentiment or news coverage are useful predictors of forward return on a firm-by-firm basis.
The second study examines the individual and joint impacts of news sentiment,
news coverage, and stock price momentum on expected return by analysing
the performance of decile portfolios formed through univariate, bivariate, and
trivariate sorts on these variables. The style of news-enhanced trading strategies
identified in the literature do not appear to be profitable over our sample
period and investment universe. However, we find some evidence that news
sentiment—if used as a screening mechanism in the short leg of momentum
strategies—can enhance risk-adjusted returns. Overall, this study provides little
evidential support for the claim that news-derived measures provide useful
conditioning information for the cross-section of returns in naive implementations.
The third study tests the economic utility of news content using a model-based
portfolio procedure in which portfolios are formed using the output of statistical
models trained over backward-looking filtrations and tested over out-of-sample periods. We find that the inclusion of news-based variables in the conditioning information set, combined with flexible statistical learning algorithms, offers only modest increase in performance beyond a traditional momentum implementation. Measures of variable importance suggest that news is secondary to size, analyst following, and momentum in relevance for predicting future return.
These findings provide little evidence that news analytics are an economically
useful data source in improving quantitative equity strategies operating over
multi-month investment horizons. Our results question the robustness and generalisability of previous findings, tested over less-stringent investment investment scopes or in higher-frequency implementations, that find news analytics to be a straight-forward means of generating excess returns.