Abstrakt
We examine the predictability of expected stock returns across horizons using machine learning. We use neural networks, and gradient boosted regression trees on the U.S. and international equity datasets.
We find that predictability of returns using neural networks models decreases with longer forecasting horizon. We also document the profitability of long-short portfolios, which were created using predictions of cumulative returns at various horizons, before and after accounting for transaction costs.
There is a trade-off between higher transaction costs connected to frequent rebalancing and greater returns on shorter horizons. However, we show that increasing the forecasting horizon while matching the rebalancing period increases risk-adjusted returns after transaction cost for the U.S.
We combine predictions of expected returns at multiple horizons using double-sorting and buy/hold spread, a turnover reducing strategy. Using double sorts significantly increases profitability on the U.S. sample.
Buy/hold spread portfolios have better risk-adjusted profitability in the U.S.