Trading strategies to exploit asset pricing anomalies hold positions for a fixed time period, usually a month. We consider early exit strategies that can close positions sooner, with possibly different exit times for different stocks in the anomaly portfolios. We show that early exit strategies, specifically ones based on machine learning techniques that use encoders to re-represent time series features, produce significant alphas and Sharpe ratios for three major anomalies, profitability, value, and momentum. Early exit strategies restore anomaly alphas in recent periods when fixed-horizon alphas have been insignificant. The results are robust to transaction costs.