Separating the signal from the noise – Financial machine learning for Twitter

Abstract

Most statistical arbitrage strategies in the academic literature soley rely on price time series. By contrast, alternative data sources are of growing importance for professional investors. We contribute to bridging this gap by assessing the price-predictive value of more than nine million tweets on intraday returns of the S&P 500 constituents. For this purpose, we design a machine learning pipeline addressing specific challenges inherent to this task. At first, we engineer domain-specific features along three categories, i.e., directional indicators, relevance indicators and meta features. Next, we leverage a random forest to extract the relationship between these features and subsequent stock returns in a low signal-to-noise setting. For performance evaluation, we run a rigorous eventbased backtesting study across all tweets and stocks. We find annualized returns of 6.4 percent and a Sharpe ratio of 2.2 after transaction costs. Finally, we illuminate the machine learning black box and unveil sources of profitability: First, results are both driven and limited by the temporal clustering of tweets, i.e., the majority of profits stem from tweets clustered closely together in time, corresponding to high-event situations. Second, the importance of included features follows an economic rationale, e.g., tweets with positive sentiment tend to yield positive returns and vice versa. Third, we find that stocks of medium market capitalization and from the consumer and technology sectors contribute most to our results, which we interpret as a trade-off between tweet coverage and tweet relevance.

Valuation Ratios, Surprises, Uncertainty or Sentiment: How Does Financial Machine Learning Predict Returns From Earnings Announcements?

“…(3) Random forests are not prone to overfitting (Breiman, 2001) and are fairly robust to noise when compared to boosting techniques (Khoshgoftaar et al, 2011). ( 4) For the specific task addressed in our paper, i.e., stock return prediction, they have empirically been found to have very good predictive performance, and perform similar or better than neural networks (Krauss et al, 2017;Gu et al, 2018;Schnaubelt et al, 2020). ( 5) Compared to neural networks and popular deep learning architectures, random forests have far less hyperparameters.…”

“…Keeping the ease of implementation in mind, we offset positions with an opposing investment in the iShares Core Total U.S. Stock Market ETF (ticker symbol ITOT, BlackRock Inc., 2020). Instead of replicating the size-decile portfolios used to calculate abnormal returns, this approach is common in the statistical arbitrage literature to achieve market-neutrality (see, for example, Avellaneda and Lee, 2010;Schnaubelt et al, 2020).…”

“…An early and directly related example is the work of Henry (2006), who applies classification and regression trees to analyze the reaction of contemporary returns to quantitative accounting information and text-based variables, and finds additional value in the verbal components of a small sample of earnings press releases. More recent empirical studies document a superior returnpredictive performance of machine learning over linear models, for example, in event studies of financial news (Ke et al, 2019;Schnaubelt et al, 2020), in applied momentum strategies (Krauss et al, 2017;Fischer and Krauss, 2018), or in evaluating large numbers of potential factors for asset pricing (Gu et al, 2018;Sak et al, 2018).…”

“…Evolution of in-sample feature importance: Next, we analyze which features are predominantly selected by the model to predict abnormal returns. FollowingKrauss et al (2017) andSchnaubelt et al (2020), we calculate feature importance as the mean decrease in impurity (MDI) for each feature during training, and average results from all study periods. Figure6depicts the evolution…”

“…For a further discussion on a timing-related attenuation of performance in event-based trading strategies, we refer the reader toSchnaubelt et al (2020).24 To exclude data errors, we have validated the returns of the respective stocks with two independent sources.…”