Working Paper
Uncertainty-Aware Portfolios
Abstract: This paper studies mean-variance portfolio choice when expected returns are estimated with heterogeneous uncertainty across assets. I model forecast uncertainty through asset-specific uncertainty radii and show that the resulting robust portfolio problem is equivalent to mean-variance optimization with an asset-specific l1 penalty. With short selling allowed, heterogeneous radii act as selective shrinkage of portfolio weights. Under long-only constraints, however, they also change portfolio composition by altering which assets clear the endogenous funding threshold. This generates two economic channels through which heterogeneous calibration can outperform a homogeneous benchmark: an active-set reallocation channel, through which capital shifts toward assets with more precise signals, and a covariance amplification channel, through which the gains from reallocation are larger when cross-asset correlation is high. I also derive finite-sample bounds linking calibration error in estimated radii to economic loss. Monte Carlo simulations and an illustrative application are consistent with these predictions.
Presentation: UMD Finance Brownbag Seminar (2026 Spring)
Generated Regressors and Inference Fragility in Two-Pass Asset Pricing
Abstract: Two-pass asset-pricing inference uses estimated factor loadings as second-pass regressors, creating a first-order inferential wedge between the benchmark with observed betas and the feasible procedure used in practice. This paper isolates that wedge and shows that its magnitude is governed by risk-premium magnitude and first-stage estimation error. Under OLS first-stage estimation, inference becomes more fragile when premia are larger and first-stage windows are shorter. In simulations, the wedge is largest in short-window designs and under misspecification. In rolling Fama–French evidence, it materially changes standard significance conclusions in characteristic-sorted portfolios. In the canonical (T1, T2) = (120,120) design, 48.3% of Shanken-significant HML windows and 57.1% of SMB windows lose significance under bootstrap calibration. Generated regressors are therefore not a technical nuisance in two-pass asset pricing, but part of the main inferential problem.
Presentation: APAD 2026
Work in Progress
When Average Fit Hides Model Failure: Tail-Robust Estimation of Stochastic Discount Factors
Price Discreteness and Quote Adjustment