What variables best predict U.S. stock market returns? In his June 2016 paper entitled “Which Variables Predict and Forecast Stock Market Returns?”, David McMillan examines the power of 25 variables to predict excess return (relative to the 3-month U.S. Treasury bill yield) of Shiller’s S&P Composite Index both in-sample and out-of-sample. He chooses variables based on connectedness to expected cash flow/dividends and risk and assigns them to five groups:
- Financial ratios: dividend-price, price-to-earnings, cyclically adjusted price-to-earnings (CAPE or P/E10), Tobin’s Q and market capitalization-to-Gross Domestic Product (GDP).
- Economic: GDP cycle, GDP acceleration (rate of change in GDP growth), consumption growth, 10-year to 3-month Treasuries term spread and inflation.
- Labor: wage growth, unemployment, natural rate of unemployment, productivity growth and labor market conditions.
- Housing: house price growth, house affordability, home ownership, housing supply and new house sales.
- Other: University of Michigan Consumer Sentiment, Purchasing Managers Index, National Financial Conditions Index, leverage and non-financial leverage.
He employs regressions to test in-sample predictive power. He then tests out-of-sample forecasts starting in 2000 using various forecast methods and accuracy measures and considering both single-variable and multi-variable models. Using the specified data series as available during 1973 through 2014, he finds that:
- From in-sample regressions spanning the full sample period, the most consistent predictors are: dividend-price ratio, price-to-earnings ratio, CAPE, GDP acceleration, inflation, the natural rate of unemployment and consumer sentiment.
- From 2000-2014 out-of-sample forecasts, models employing multiple predictive variables almost always outperform all those using only one.
In summary, evidence suggests that investors should combine several different kinds of predictors when estimating expected stock market returns.
Cautions regarding findings include:
- Some predictors have revision bias (current historical series do not match as-released data), thereby confounding both in-sample and out-of-sample tests. See, for example, “Productivity and the Stock Market”.
- Shiller’s S&P Composite Index data are “blurry” since monthly values are averages of daily values.
- The sample is very short for analysis of CAPE, consisting of only about four independent 10-year average earnings measurement intervals.
- Testing many variables, methods and performance metrics on the same stock market return data introduces snooping bias, thereby overstating expectations for the best (luckiest) predictors.
- The author does not address economic import of findings via any market timing strategies.
See also:
