Can investors effectively use firm characteristics to screen European stocks? In their August 2019 paper entitled “Predictability and the Cross-Section of Expected Returns: Evidence from the European Stock Market”, Wolfgang Drobetz, Rebekka Haller, Christian Jasperneite and Tizian Otto examine the power of 22 firm characteristics to predict stock returns individually and jointly. They assume market-based characteristics are available immediately and accounting-based characteristics are available four months after firm fiscal year end. For multi-characteristic predictions, they consider 5-characteristic, 8-characteristic and 22-characteristic models. For regression-based forecasts, they use either 10-year rolling or inception-to-date monthly inputs. For economic tests, they form equal-weighted or value-weighted portfolios that are each month long (short) the tenth, or decile, of stocks with the the highest (lowest) expected next-month returns based on 22-characteristic regression outputs. To estimate net performance, they apply one-way trading frictions of 0.57%. Using groomed monthly data for all firms in the STOXX Europe 600 index during January 2003 through December 2018, they find that:
- On average, the number of firms with sufficient data for testing is 458, ranging from 310 in January 2003 to 543 in December 2018. Average (equal-weighted) monthly gross excess return is 1.09% with standard deviation 7.12%.
- Regarding multi-characteristic model regressions:
- In general, including more characteristics in regressions improves model performance.
- In-sample, abilities to explain month-ahead stock returns range from adjusted R-squared 0.07 for a 5-characteristic model to 0.15 for a 22-characteristic model. Explanatory powers tend to weaken over time.
- Out-of-sample, abilities to forecast month-ahead stock returns range from adjusted R-squared 0.014 for a 5-characteristic model to 0.016 for a 22-characteristic model. These values, while small, are statistically significant.
- For the 22-characteristic model, experimental exclusion of individual characteristics discovers only two significant findings: (1) excluding size hurts model performance; and, (2) excluding long-term reversal improves model performance.
- Regarding portfolios formed from extreme deciles of expected monthly returns based on 22- characteristic monthly forecast regressions:
- The long-short equal-weighted (value-weighted) portfolio generates average monthly gross return 1.84% (1.13%), with annualized gross Sharpe ratio 2.26 (1.02). Monthly 5-factor (market, size, book-to-market, investment, profitability) gross alphas are similar to monthly returns.
- A long-only strategy that each month holds the equal-weighted (value-weighted) decile of stocks with the highest expected returns generates average monthly gross excess return 2.15% (1.35%), with annualized gross Sharpe ratio 1.23 (0.74).
- Deducting 0.57% one-way portfolio turnover frictions from long-only monthly excess returns results in equal-weighted (value-weighted) average annualized net excess return 18.0% (7.5%), with annualized net Sharpe ratio to 0.85 (0.37) and maximum drawdown -70% (-75%). For the value-weighted market, annualized excess return is 5.6%, with Sharpe ratio 0.40 and maximum drawdown -56%.
- Results are generally consistent for monthly, quarterly and annual return forecast intervals.
In summary, evidence indicates that investors in European stocks may be able to boost performance by screening jointly for about 20 firm characteristics.
Cautions regarding findings include:
- As footnoted in the paper, the authors use an inception-to-date regression during the first 10 years of the sample for the monthly “10-year rolling” regressions because the entire sample has duration only 15 years. This approach highlights the fact that the sample period is extremely short in terms of the modeling approach.
- Reported drawdowns of extreme decile portfolios are so deep that many investors would not tolerate them.
- Value-weighted extreme decile portfolios appear substantially less attractive than equal-weighted counterparts (and not clearly preferable to the value-weighted market on a net basis), in some conflict with the discussion in the paper. In mitigation, the authors argue that estimated trading frictions are very conservative.
- Equal-weighted portfolios are likely more costly to trade than value-weighted counterparts due to substantial tilt toward less liquid stocks.
- Testing many characteristics and combinations of characteristics on the same sample introduces data snooping bias. Selecting the best combination and regression method for detailed testing carries such bias into subsequent findings. Selecting characteristics for testing based on prior studies incorporates any snooping bias impounded in those studies.
- Data collection/processing and portfolio maintenance as described are beyond the reach of most investors, who would bear fees for delegating to a fund manager.
