Equity Premium Prediction by Sparse Pooling of Parsimonious State-Dependent Models

A large set of macroeconomic variables have been suggested as equity risk premium predictors in the literature. This paper proposes a forecasting approach for the equity risk premium with two novel features. First, individual month-ahead forecasts are obtained from parsimonious threshold regression models that exploit from one to at most three macroeconomic predictors jointly, and a binary technical indicator as an observable state variable to accommodate expansion versus recession predictability states. Second, a sparse combination of those forecasts is carried out through a robust predictability test. A comprehensive out-of-sample forecast evaluation exercise based on statistical criteria and asset-allocation criteria demonstrates that both features of the proposed approach enable gains versus existing forecasting techniques. However, the state-dependent aspect of the forecasts delivers larger improvements in forecast accuracy that the sparse combination aspect. The results are robust to sub-period analyses, expanding versus rolling estimation windows, and different investors' risk aversion levels