Do the methods and assumptions used in studies of the power of variables to predict differences in future returns across stocks accurately represent implementable trading strategies? In his December 2009 paper entitled “Economic and Statistical Properties of Implementable Trading Strategies”, Andrew Christie assesses the realism of widely used portfolio-level tests for anomalous cross-sectional stock returns. Using analysis and (as an example) the results from some past portfolio studies on the predictive power of standardized unexpected earnings, he concludes that:
- Implementable trading strategies must be ex ante (based on real-time expected values) and net (incorporating all search costs and trading frictions).
- Suppose an investor selects stocks for a portfolio based on the distribution of expected returns (across stocks) associated with historical values of some predictive variable. In general, elevating the information content of the portfolio (by moving further into the tails of the distribution) also increases the proportion of high-mean, high variance stocks in it.
- As investor willingness to accept risk decreases, the number of stocks selected tends to increase. In other words, risk-averse investors seek to exploit broadly diversified differences rather than narrowly diversified extremes.
- As search costs and trading frictions rise, the number of stocks selected tends to decrease because the part of the return distribution that is net profitable shrinks.
- Also, securities carrying higher information content per the predictive variable tend to have higher trading frictions (smaller firms, stocks with lower price and lower liquidity). Traders must perform real-time analysis on the net return by security. Accounting for trading frictions by security affects portfolio selection.
- Zero-cost (long-short hedge) portfolios do not really have zero cost because of the margin requirements of shorting and portfolio rebalancing costs. The old (pre-July 2007) uptick rule also distorts actual frictions for shorting in historical data. In other words, the long and short sides of such a portfolio do not actually offset, and zero-cost portfolios have a positive return bias.
- Data snooping bias is substantial, and publishing bias (whereby null outcomes do not get submitted or selected for publication) may suppress the detectable intensity of actual snooping.
- The above considerations suggest that trading strategies used in well-known studies of standardized unexpected earnings do not earn abnormal returns based on real-time publicly available information.
In summary, investors may want to ensure that they base trading strategies on real-time expectations net of search costs and comprehensive trading frictions, with a substantial margin to accommodate data snooping bias.
