Revisiting Momentum Through Dynamic Kernel Learning

This paper revisits momentum through a dynamic kernel framework that links return continuation to the speed at which return‐relevant shocks propagate through prices. Using nearly 60 years of US equity data, we estimate nonnegative and geometric kernels that govern how past returns are weighted over time. The estimated decay parameter rises sharply after 2000, indicating faster dissipation of return persistence and a shorter effective horizon for continuation, coinciding with the documented weakening of traditional 12–1 momentum. In out‐of‐sample tests, the structured kernel models match or outperform common machine‐learning benchmarks while retaining economic interpretability, and these gains remain robust after controlling for volatility timing and trading frictions. The framework provides a parsimonious tool for tracking time variation in return propagation and for adapting momentum implementations to changing market conditions.