Changepoint detection in the cross-section of stock returns
使用多种无监督学习方法研究股票日收益横截面中的均值、方差和分布变点，发现变点普遍存在且GARCH模型无法解释方差变点，对长期因子模型的可解释性构成挑战。
Abstract We investigate changepoints in the cross-section of stock returns using an ensemble of dedicated unsupervised learning methods. Our large-scale study reveals a sustained incidence of changepoints in the mean, variance, and distribution of daily returns. This finding is robust to the choice of the changepoint detection method. We also find that a six-factor empirical asset pricing model partially explained these results until the mid-1990s, but its explanatory power has weakened considerably. Moreover, GARCH models do not account for changepoints in the variance of the residuals. This striking finding indicates that conditional heteroskedasticity is not connected to changepoints, at least not in a way standard econometric models capture. A further study on monthly data confirms that changepoints are a robust feature of the cross-section. Predictably, the number of detected changepoints is smaller than with daily data, yet these changepoints persist across methods, especially for variance and distribution. Therefore, changepoints undermine the interpretability of constant-parameter Fama–French-style factor models on long samples: least-squares estimates collapse multiple regimes into a single set of parameters, and residual-based inference weakens. Accordingly, changepoint detection should be incorporated as a standard component of model validation and recalibration.