Our Media-based Stock Prediction Model, Continued Outperformance

We built a 30-day U.S. stock predictive model based on Media Sentiment using a simple Machine Learning algorithm - updated returns in the below image. Humans, even in hyper-competitive environments like the US stock market, respond predictably to news and social media content.

The model was put in production in Aug 2019 and launched commercially on Jan 2020 in partnership with the StarMine team at LSEG Data & Analytics.

The model ranks stocks (1 to 100) based on their likelihood of outperformance using a ML algo applied to only our LSEG MarketPsych Analytics news and social media-derived data. The blue line in the chart below depicts the spread between the top-ranked decile (10%) of stocks and the bottom-ranked decile - the absolute return portfolio. The gray-shaded regions are the 2-year training periods.

The model performs similarly regardless of market cap. It is orthogonal to other StarMine models. It has about 30% monthly turnover, with no transaction costs assumed. Longer term models also work well with media data (we've tested out to 3 year forecasts).