Out-of-sample equity premium predictability: An EMD-denoising based model

The poor out-of-sample forecasting performance of the stock returns of various predictors has been widely confirmed in the literature, which casts doubt on the reliability of stock-return predictability. However, the reliability of return predictability is closely related to the noise contained in the data. In this study, we design a new method to address the noise in the framework of empirical mode decomposition. The EMD method provides an efficient return decomposition, and based on which we selectively remove high-frequency components that are more likely to be contaminated by outliers. Our new model delivers statistically and economically significant out-of-sample gains relative to the historical average. The predictive ability mainly originates from the business-cycle risk and survives a series of robustness tests. •We introduce the EMD method and economic constraints to denoise financial time-series data.•Our method provides economic meaning and may serve as a better alternative to the EMD literature in Computer Science.•We make one-step-ahead forecasts for stock market returns (S&P500 index) over the period ranging from January 1947 to December 2020.•Our EMD method delivers both statistically and economically significant gains in forecasting performance.