Tzikas, Candès, Hastie, Boyd, Kochenderfer & Kahn (2026) — Transient Statistical Factors for Risk Models
Authors: A. E. Tzikas, E. J. Candès, T. Hastie, S. P. Boyd, M. J. Kochenderfer, R. N. Kahn. Title: Enhancing a Risk Model by Adding Transient Statistical Factors. Venue: arXiv preprint arXiv:2605.12977 (2026).
What the paper is
A method to augment an existing low-rank-plus-diagonal factor risk model with k new transient statistical factors, fit by weighted maximum likelihood and EM. Keeps base factor loadings fixed; improves sample-out-of-sample R^2, residual R^2, log-likelihood, and whitened-return correlation on Barra USE4 + 870 US large-caps.
Why the author list matters
This is a rare combination of three communities:
- Stanford convex optimization — Boyd, Candès.
- High-dimensional statistics / sparse methods — Hastie, Candès.
- BlackRock systematic investing — Kochenderfer, Kahn (the latter is the long-time head of BlackRock's risk model group).
So the paper is at once theoretically clean (KL/MLE equivalence, EM convergence, missing-data marginalization) and operationally grounded (Barra USE4 as the base, EWMA half-life as a knob, production-engineering failure modes called out).