Is information technology moving in on qualitative event trading just as it has high-frequency quantitative algorithm trading? In the October 2011 version of their paper entitled “Event Driven Trading and the ‘New News'”, David Leinweber and Jacob Sisk examine the trading acumen of a model (set of filters) trained to exploit Thomson Reuters News Analytics metadata (sentiment tone, stock relevance and novelty as measured by link counts). Their portfolio simulation approach: (1) is restricted to the technology, industrials, health care, financials and basic materials sectors; (2) assumes an extreme sentiment day for a stock has at least four novel news items (prior to 3:30PM in New York) and is among the top 5% of average daily positive or negative events; (3) makes portfolio changes at market close; (4) holds positions for 20 days, subject to a 5% stop-loss rule and a 20% take-profit rule; (5) constrains any one position to 15% of portfolio value; and, (6) assumes round-trip trading friction of 0.25%. Using news metadata for the S&P 1500 and associated stock returns during 2003 through 2009 for in-sample testing and the first three quarters of 2010 for out-of-sample testing, they find that:
- In-sample testing indicates that:
- There is a trade-off in filter settings between number of signals generated and signal exploitability.
- Negative sentiment signals are more exploitable than positive signals.
- Signals for small and medium capitalization stocks are stronger than those for large capitalization stocks.
- Returns are volatile, with maximum drawdown about 60%. Mean monthly return is 1.7%, with 52% of months profitable.
- The filter models starts producing alpha in 2007 when the Thomson Reuters News Analytics metadata increases dramatically in terms of breadth, depth and volume.
- The largest alpha events cluster, such that many of the “alpha spikes” derive from short positions during the financial crisis when it was difficult or impossible to take these positions.
- During the nine-month out-of-sample test, a portfolio driven exclusively by the news metadata filter model beats the S&P 500 Index by a net 11.5% (see the chart below).
The following chart, taken from the paper, compares the out-of-sample cumulative trading performance of the news metadata filter model to that of the S&P 500 Index (apparently excluding dividends) during the first three quarters of 2010. During this period, the model generates a positive return and beats the index by 11.5%.

In summary, evidence indicates that traders may be able to beat the stock market by systematically filtering a broad and deep source of news metadata to identify sentiment extremes for individual stocks.
Cautions regarding findings include:
- The out-of-sample test period is short in terms of variety of market conditions.
- Tests apparently do not allocate any cost of using the Thomson Reuters News Analytics service to trading friction. This cost would reduce reported returns.
- Since smaller capitalization stocks apparently drive profitability and test portfolio allocation rules are not be based on market capitalization, the S&P 500 Index may not be an appropriate benchmark. For example, the total return for iShares S&P 1500 Index (ISI) during the first three quarters of 2010 is 4.2%, compared to 3.5% for S&P 500 SPDR (SPY). And, the total return for Rydex S&P 500 Equal Weight (RSP) during the first three quarters of 2010 is 8.1%.
- The assumed level of trading friction may be optimistic for many traders, especially during market crises or narrower crises specific to traded stocks.
- Market adaptation to widespread use of news metadata is plausible.
Frequently Asked Questions
What is a news sentiment filter model in stock trading?
A news sentiment filter model is a set of rules that screens news analytics metadata, such as sentiment tone, stock relevance and novelty, to identify stocks with extreme positive or negative sentiment for trading. In the Leinweber and Sisk study, an extreme sentiment day required at least four novel news items before 3:30PM New York time and a stock ranking in the top 5% of average daily positive or negative events.
Can trading on news sentiment extremes beat the S&P 500?
Trading on news sentiment extremes beat the S&P 500 in one short test. During the first three quarters of 2010 (out-of-sample), a portfolio driven only by the news metadata filter model outperformed the S&P 500 Index by a net 11.5%. The result comes from a nine-month test period and may not hold across varied market conditions.
How is a news analytics trading strategy tested using in-sample and out-of-sample data?
The news analytics strategy was trained on Thomson Reuters News Analytics metadata for the S&P 1500 during 2003 through 2009 (in-sample) and tested on the first three quarters of 2010 (out-of-sample). The simulation changes positions at market close, holds them 20 days with a 5% stop-loss and 20% take-profit, caps any position at 15% of portfolio value, and assumes 0.25% round-trip trading friction. It is limited to the technology, industrials, health care, financials and basic materials sectors.
Is the S&P 500 the right benchmark for a news sentiment strategy that favors small-cap stocks?
The S&P 500 may not be an appropriate benchmark for a news sentiment strategy driven by small and medium capitalization stocks, since the portfolio rules are not based on market capitalization. In the first three quarters of 2010, total return was 4.2% for iShares S&P 1500 Index (ISI), 3.5% for S&P 500 SPDR (SPY), and 8.1% for Rydex S&P 500 Equal Weight (RSP). Against the equal-weight alternative, the 11.5% outperformance claim looks less dramatic.
What are the risks and limitations of trading on news sentiment signals?
Trading on news sentiment signals carries several limitations: the out-of-sample test was short and covered limited market conditions, and in-sample returns were volatile with a maximum drawdown of about 60%. The tests apparently exclude the cost of the Thomson Reuters News Analytics service, and the 0.25% trading friction may be optimistic, especially during crises. Many of the largest alpha spikes came from short positions during the financial crisis, when such positions were difficult or impossible to take, and wider use of news metadata could erode the edge.