News Sentiment Analysis

News and Media Sentiment Analysis for Trading

The intersection of news, sentiment analysis and trading strategies has been considered for a number of years. Work ranging from Tetlock et al (2005) to the more recent efforts by Leinweber et al (slides) (paper) have established relationships between sentiment on publicly available news information and future changes in asset value.

We recently implemented a sentiment metric on entities (e.g. General Electric, Microsoft, etc.) and events (e.g. IPO Tesla, Product Launch Pfizer) that we process and decided to make an initial assessment of this metric in the context of trading strategies. Are there relationships between our sentiment metric and is there evidence that these can be predictive for asset prices. With these questions in mind, we executed a project to take a closer look.

We developed and tested our sentiment metric based on business related documents so it is generally tuned for company/financials related news. The sentiment metric is instantaneously measured on all entities and events in documents that we process and higher values of the metric correspond to higher levels of positive sentiment.

We examined the sentiment around a collection of earnings related content in the system and analyzed average sentiment measures around earnings calls. Earnings calls obviously have a large impact on asset values and one would expect that earnings calls surrounded by positive sentiment would be followed by a rise in price and calls surrounded by negative sentiment a fall.

Figure 1: Correspondence between calculated sentiment and the Close of the S&P500

First, we simply calculated our average sentiment measure over time for S&P500 companies compared to the close of the S&P 500. As you can see in the above graph (made using Spotfire Professional), our sentiment metric tracked the S&P fairly well although it does seem to lag the index for some of the time. This was actually very encouraging. We didn't include any pricing data in the development or testing of our sentiment metric so seeing this close a relationship suggested that we were measuring a real phenomenon with our metric.

Figure 2: Relationship between calculated sentiment on earnings call presentations and two selected companies

We next drilled down and took a look at a few individual companies and as you can see above, saw reasonable correlation with some companies such as Google and less with other companies such as Pfizer. Our final step in this first analysis was to take a deeper look at longer term phenomena. Specifically, we looked at the difference in sentiment scores between quarters and whether this value was useful in predicting asset values. We implemented a paper trading strategy to evaluate our sentiment signal based on earnings calls for S&P 500 companies back to 2006. We "bought" a stock the morning after the call if we saw an increase in the call related sentiment between quarters and "shorted" a stock accordingly if the sentiment fell. We then held on to the position until the next earnings call event for that company and then either maintained the position or reversed it based on the difference between the sentiment aggregated for the new quarter's call and the previous one. We looked at this on a return basis, rebalanced the portfolio daily to keep the positions in each company to an identical amount and averaged the return across the companies we were invested in at any specific point in time. In this approach, we ignored trading costs.

Figure 3: Returns from sentiment based trading strategy compared to S&P500 performance


The above plot illustrates the returns our approach yielded as well as the returns from the S&P500 over the three year period from April, 2007 to Mar, 2010. We started in 2007 because by this time we had sufficient historical content associated with the vast majority of the S&P500 companies to perform our calculations. Our total return for the this period was 15% (s.d. of daily return was 0.7%) compared to -18% (s.d. 1.9%) for the S&P500. We yielded 24/36 profitable months as opposed to the 18/36 profitable months observed with the S&P500. Our approach tends to under-perform the market when the market is rising, but over-perform the market, sometimes significantly, when the market is declining.

Taking a closer look at each of the three covered years independently confirms our view that the approach is generally doing well in market downturns, but is missing the upturns, suggesting that negative sentiment may be more telling in this context then positive sentiment.

Figure 4. Three 12 month segments of strategy performance vs. S&P500

Since we didn't account for trading costs and since the approach under-performs a rising market, we aren't done working with this signal. However, we were encouraged to see that the overall correlation we saw in Figure 1 can be used to derive a reasonable starting point for a trading strategy that outperforms the market overall for this three year period. Further investigation of this signal will include:

1. Subdividing by Industry/Market Cap - Leinweber illustrates that sentiment can have a different impact based on the market cap and industry of the set of assets evaluated. Perhaps this approach will work with technology companies like Google, but not Pharmaceutical companies like Pfizer.
   
2. Modifying position holding times - With essentially a three month holding period for each position, we are at risk for any other trends in the stock to change the fundamental behavior. Depending on when in the time the gains typically come with the stock, there may be room for optimizing with shorter holding times
   
3. Incorporating stop loss and profit taking rules - Another strategy to reduce holding times is to base them on stock performance criteria
   
4. Improved use of the sentiment scores. Currently we are only examining the change in sentiment direction, but not considering the magnitude of that change or comparing the magnitudes across companies. Restricting to the larger changes, either in relative or absolute terms may improve performance.
5. Incorporating other content. We calculate other metrics such as momentum and a hedging score. Using these scores could help.
6. Only include documents/events that are decisively not about price change that has already happened (e.g. avoid documents saying "Pfizer was down 3% in late trading yesterday").
   

Our efforts here support the current research in the area, that there is promise of finding alpha from the linguistic analysis of publicly available content. The challenge is to find and implement the optimal approaches.