Breakpoint estimation
Financial and economic time series change regime. After a crisis, a policy change, or a shift in market structure, the way assets move together is no longer the same. A basic statistical question is to locate the moment of change, the breakpoint, from the data alone. The quasi-maximum likelihood (QML) estimator does this by splitting the series at a candidate time \(\nu\) and comparing the log-determinants of the sample covariances before and after the split; the estimated breakpoint is the split that minimizes this criterion.
With Hubeyb Gurdogan we prove that the QML estimator is consistent in a model-free setting: no stochastic data-generating model, no independence, and no distributional assumptions. What decides consistency is a signal-to-noise ratio, the contrast between the two regimes compared with the fluctuations within each regime. If this ratio exceeds an explicit threshold, the estimated breakpoint converges to the true one. The threshold is sharp: signals that violate it by an arbitrarily small margin can push the minimizer to the wrong location. We then apply the framework to pervasive factor models whose dimension grows with the sample size, which extends earlier results of Duan, Bai and Han to error terms that need not be independent, idiosyncratic, or random.
Natural next steps are several breakpoints, heterogeneous noise, rates of convergence, and a comparison with computationally cheaper estimators that track angles or eigenvalues instead of full log-determinants. The question fits the theme of this page: when does a noisy, high-dimensional system reveal a reliable macroscopic signal?
Further reading.
- A sharp signal-to-noise threshold for quasi-maximum likelihood breakpoint estimation, with H. Gurdogan. arXiv:2609.12271, 2026.