A hot topic in financial modeling is using news to augment trading strategies. If you’ve been following these recent developments, you may have heard about Thomson Reuters News Analytics (TRNA), an automated real-time news analysis system that uses natural language processing to analyze news. In this article, we’ll describe two ways of using TRNA that show promise as part of a profitable trading strategy. Both demonstrate how combining TRNA with a StarMine stock-ranking signal can provide better results than either signal alone.
TRNA and Short Term Price Momentum
As you might guess, good news is bullish and bad news is bearish. We can amplify this effect by combining TRNA with the fast-moving StarMine Short Term Price Momentum (PriceMoST) factor. PriceMoST predicts near-term returns based on reversion of recent price movement, and scores stocks on a 1-100 scale (100 = most bullish).
Below is an event study covering the S&P 1500 index from 2003-2013. For each company, we start with all Reuters news in the preceding 8 hours, filter on TRNA relevance > 0.8, and identify an “event” if average TRNA net sentiment is > 0.75 (positive events) or < -0.75 (negative events).
Positive and negative events yield on average a 25 basis point spread over 10 days (lighter-toned lines).
The darker lines show positive events including only stocks with PriceMoST > 75 (blue) and negative events with PriceMoST < 25 (red). On average, that spread is 55 basis points, yielding an impressive 30 basis-point improvement over news sentiment alone. TRNA and Predicted Earnings Surprise
StarMine has been rating analysts on their estimate accuracy for over a decade. Some analysts are consistently more accurate forecasters than others. StarMine captures this phenomenon with its proven SmartEstimate and Predicted Surprise measures. The StarMine SmartEstimate re-weights the consensus estimate, emphasizing more recent estimates and top-rated analysts. Predicted Surprise is the percent difference between the SmartEstimate and the I/B/E/S consensus estimate. When SmartEstimates diverge significantly from consensus, it serves as a leading indicator of the direction of future revisions and/or surprises. In aggregate, a Predicted Surprise of 2% or more correctly anticipates the direction of actual earnings surprises roughly 70% of the time.
To show TRNA’s flexibility, we construct here a different signal than in the Price Momentum example. Let’s look at what happens around earnings announcements when we combine news with Predicted Surprise.
We studied S&P 1500 earnings announcements between 2006 and 2011 using average net news sentiment over the preceding 7 days to rank securities on a 1-100 scale (100 = most positive sentiment).
The heat map above shows average excess returns from market close 1 day before the announcement to market close 5 days after. The diagram is divided into 15 bins depending on Predicted Surprise and pre-announcement TRNA rank.
A trading strategy based on combining TRNA and Predicted Surprise might go long when the signals are both bullish (upper right-hand corner) and short when they are both bearish (lower left-hand). The spread between these two bins is a whopping 110 basis points – an unusually large return for a short-term news strategy, and enough to easily overcome transaction costs.
You could also benefit from TRNA when following positive Predicted Surprises alone. The heat map’s top row shows that excess returns when TRNA and Predicted Surprise agree are significantly higher (again 110 basis points) than when they disagree.
Let’s dive into an example: Harley-Davidson Inc. (HOG) and Barnes and Noble Inc. (BKS). In January 2011, both had very positive Predicted Surprises, but HOG had positive recent news while BKS had negative recent news. How did the market react to their earnings announcements?
After reporting and beating estimates (-$0.18 actual Q4 2010 earnings versus -$0.26 consensus), HOG (blue) beat the market by about 500 basis points over the subsequent 5 days.
BKS reported an EPS miss ($1.00 actual versus $1.13 consensus), and its subsequent 5-day excess return (red) was about -2000 basis points.
In both cases, our combination signal predicted the direction of subsequent excess stock returns.
These examples demonstrate the value of combining news sentiment with other fundamental factors, and they show that the combined benefits persist long enough to enhance many traditional (non-high frequency) trading strategies.
 Stauth, J., and Bonne, G. “SmartEstimates and the Predicted Surprise: Construction and Accuracy”, StarMine white paper, 2009
Receive stories like this to your inbox as they are published. Subscribe here and follow us @Alpha_Now on Twitter. If you are looking to access Thomson Reuters data or analytics, register for a free trial.