A frequency-domain alternative to long-horizon regressions with application to return predictability

• Frequency-domain method for testing for long-run predictability in stock returns
• The method is compared to standard approaches: long-horizon and simple regressions.
• The frequency-domain method always outperforms long-horizon regressions.
• The comparison with simple regressions depends on short-run dynamics.
• We find evidence of return predictability even with subsampled confidence intervals.

This paper aims at improved accuracy in testing for long-run predictability in noisy series, such as stock market returns. Long-horizon regressions have previously been the dominant approach in this area. We suggest an alternative method that yields more accurate results. We find evidence of predictability in S&P 500 returns even when the confidence intervals are constructed using model-free methods based on subsampling.