Can investors exploit economic data for monthly stock market timing? In their September 2015 paper entitled “Getting the Most Out of Macroeconomic Information for Predicting Excess Stock Returns”, Cem Cakmaklı and Dick van Dijk test whether a model employing 118 economic variables improves prediction of monthly U.S. stock market (S&P 500 Index) excess returns based on conventional valuation ratios (dividend yield and price-earnings ratio) and interest rate indicators (risk-free rate, change in risk-free rate and credit spread). Excess return means above the risk-free rate. They each month apply principal component analysis to distill from the 118 economic variables (or from subsets of these variables with the most individual power to predict S&P 500 Index returns) a small group of independent predictive factors. They then regress next-month S&P 500 Index excess returns linearly on these factors and conventional valuation ratios/interest rate indicators over a rolling 10-year historical window to generate excess return predictions. They measure effectiveness of the economic inputs in two ways:

- Directional accuracy of forecasts (proportion of forecasts that accurately predict the sign of next-month excess returns).
- Explicit economic value of forecasts via mean-variance optimal stocks-cash investment strategies that each month range from 200% long to 100% short the stock index depending on monthly excess return predictions as specified and monthly volatility predictions based on daily index returns over the past month, with transaction costs of 0.0%, 0.1% or 0.3%.

Using monthly values of the 118 economic variables (lagged one month to assure availability), conventional ratios/indicators and monthly and daily S&P 500 Index levels during January 1967 through December 2014, *they find that:*

- Over the entire sample period, monthly excess return prediction directional accuracies are:
- 58.5% using only conventional valuation ratios and interest rate indicators.
- 55.2% using only the three economic variables that are the best individual predictors.
- 62.1% augmenting conventional predictors with the distilled full set of 118 economic variables.
- 57.5%-58.5% augmenting conventional predictors with different ways of pre-screening and then distilling economic variables.
- 58.5% using only the historical average excess return (almost always predicting positive values).

- Net Sharpe ratios for a long-only/no leverage mean-variance optimal monthly allocation strategy with 0.3% transaction costs are:
- 0.53 for buying and holding the market.
- 0.49 using only conventional valuation ratios and interest rate indicators.
- 0.41 using only the three economic variables that are the best individual predictors.
- 0.63 augmenting conventional predictors with the distilled full set of 118 economic variables.
- 0.46-0.73 augmenting conventional predictors with different ways of pre-screening and then distilling economic variables.
- 0.36 using only the historical average excess return.

- An important aspect of the attractiveness of augmenting conventional predictors with economic factors is the failure of the former during the 1990s.

In summary, *evidence suggests that augmenting conventional valuation ratios and interest rate indicators with a large set of lagged economic indicators (distilled via principal component analysis) may modestly improve prediction of the equity risk premium at a monthly horizon.*

Cautions regarding findings include:

- It is not clear whether S&P 500 Index returns include dividends, which may materially affect a contest between an active strategy and buy-and-hold. The corresponding author did not respond to a request for clarification.
- The medium (0.1%) and high (0.3%) levels of transaction costs used for investment strategy tests are very low for much of the sample period (see “Trading Frictions Over the Long Run”). Prediction exploitation results may therefore be very optimistic. The corresponding author did not respond to an inquiry about assumed levels.
- Strategy implementation costs apparently do not include any net costs of shorting or leverage when used.
- The study does not account for costs of timely data collection and processing, which is beyond the reach of most investors. Delegating these efforts to an advisor or fund manager would involve fees.
- Testing several model variations on the same data introduces snooping bias, such that the best-performing variation overstates expectations for predictive power.