@vince.quant: What if having less data actually made your portfolio better? Classical portfolio theory says this should fail. When N > T, the sample covariance matrix becomes singular and you can construct portfolios with zero in-sample variance. Most people would call this overfitting. But a recent preprint by Chang, Ding, Shi and Zhang (2026) shows something more subtle. Using a simple estimator called Ridgelet, which adds a tiny fixed perturbation to the covariance matrix, they show that out-of-sample risk follows a double descent curve. Risk increases as you approach the interpolation threshold N = T, but once you move deep into the overparameterized regime N >> T, it decreases again. The intuition: when many portfolios perfectly fit the data, the estimator selects the minimum L2-norm solution, which can generalize well. The same phenomenon that drives modern overparameterized neural networks shows up in portfolio construction. This is a preprint, not yet peer-reviewed. Limitations: -> The theory relies on factor model assumptions and random matrix asymptotics -> The "good" regime requires N >> T, not just slightly above -> Performance depends on the covariance structure -> A naive pseudoinverse approach fails badly out-of-sample Paper -> arXiv:2602.19462 #finance #quant #trading #algotrading #stocks

vince.quant

Open In TikTok:

Region: FR

Tuesday 28 April 2026 23:10:35 GMT