Shift-Aware Gaussian-Supremum Validation for Wasserstein-DRO CVaR Portfolios

Derek Long · 2025

Read Shift-Aware Gaussian-Supremum Validation for Wasserstein-DRO CVaR Portfolios by Derek Long (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2512-16748.txt.

Introduction

Derek Long studies how to pick stock portfolios that keep tail risk in check when markets shift. The task uses a CVaR limit, a cap on losses in the worst cases. A robust method guards the cap. It hedges against a ball of nearby chance models.

Picking the ball size is the snag. Too big is timid. Too small can break under regime change. Long gives a two-phase check. It first solves the robust task over a grid of sizes. It then reweights a later data fold to mimic the target market. It also builds a confidence band that respects time links. Any portfolio that passes meets the risk cap with set odds. Tests show better return-risk trade-offs than a naive baseline.

Catalog record

CREATORDerek Long
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE28 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2512-16748

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Shift-Aware Gaussian-Supremum Validation for Wasserstein-DRO CVaR Portfolios Derek Long Abstract We study portfolio selection with a Conditional Value-at-Risk (CVaR) constraint under distribution shift and serial dependence. While Wasserstein distributionally robust optimization (DRO) offers tractable protection via an ambiguity ball around empirical data, choosing the ball radius is delicate: large radii are conservative, small radii risk violation under regime change. We propose a shift-aware Gaussian-supremum (GS) validation framework for Wasserstein-DRO CVaR portfolios, building on the work by Lam and Qian (2019). Phase I of the framework generates a candidate path by solving the exact reformulation of the robust CVaR constraint over a grid of Wasserstein radii. Phase II of the framework learns a target deployment law Q Q by density-ratio reweighting of a time-ordered validation fold, computes weighted CVaR estimates, and calibrates a simultaneous upper confidence band via a block multiplier bootstrap to account for dependence. We select the least conservative feasible portfolio (or abstain if the effective sample size collapses). Theoretically, we extend the normalized GS validator to non-i.i.d. financial data: under weak dependence and regularity of the weighted scores, any portfolio passing our validator satisfies the CVaR limit under Q Q with probability at least 1 − β 1-\beta ; the Wasserstein term contributes a deterministic margin ( δ / α ) ​ ‖ x ‖ ∗ (\delta/\alpha)\|x\|_{*} . Empirical results indicate improved return–risk trade-offs versus the naive baseline. \workshoptitle Generative AI in Finance 1 Introduction Portfolio optimization under uncertainty requires robust risk management, especially for extreme losses. A common risk metric is the Conditional Value-at-Risk (CVaR), which measures the average loss in the worst \alpha fraction of cases. CVaR constraints are used to cap tail risk and prevent systemic crises by ensuring portfolio losses remain below a threshold with high confidence. In practice, data-driven portfolio selection with a CVaR constraint faces significant challenges: the true distribution of returns is unknown and may shift over time (regime shifts), and asset returns are often dependent (serially correlated and cross-correlated), violating i.i.d. assumptions. Classical robust optimization or distributionally robust optimization (DRO) approaches account for some distributional uncertainty by considering worst-case distributions in an ambiguity set (e.g. a Wasserstein ball around the empirical distribution). However, these methods can be sensitive to unanticipated distribution shifts: if market conditions change (e.g. a volatility regime change), a previously feasible portfolio may violate the CVaR limit. Moreover, dependent data reduces effective sample size and can lead to mis-estimated risk. This motivates the need for a shift-aware validation procedure that guarantees CVaR constraints under distribution shifts and dependence. We focus on the CVaR-constrained portfolio problem with d assets. Let x\in\mathbb{R}^{d} be allocations, \xi\in\mathbb{R}^{d} asset returns, F the (unknown) data-generating law, and c=-\mathbb{E}_{F}[\xi]\in\mathbb{R}^{d} (expected losses). The classical formulation is \displaystyle\underset{x}{\mathrm{min}} \displaystyle c^{\top}x (1) \displaystyle\mathrm{s.t.} \displaystyle\inf_{t\in\mathbb{R}}\Big\{t+\tfrac{1}{\alpha}\mathbb{E}_{F}[\max\{-\xi^{\top}x-t,\,0\}]\Big\}\leq\gamma, \displaystyle\sum_{l=1}^{d}x_{l}=1,\qquad x_{l}\geq 0~\forall l\in\{1,\ldots,d\}. To guard against distributional uncertainty, we adopt a Wasserstein DRO version centered at an empirical law \hat{F} with radius \delta , so that the robust CVaR constraint becomes \sup_{Q\in\mathcal{B}(\hat{F},\delta)}\inf_{t\in\mathbb{R}}\Big\{t+\tfrac{1}{\alpha}\mathbb{E}_{Q}[\max\{-\xi^{\top}x-t,\,0\}]\Big\}\leq\gamma.\\ (2) For W_{1} ambiguity balls, the robust CVaR constraint admits an exact, tractable reformulation (cf. Cor. 5.1 in [ 9 ] ): Z=\left\{x\in\mathbb{R}^{d}:\begin{array}[]{l}\delta v+\frac{1}{n}\sum_{i=1}^{n}z_{i}\leq\alpha r,\\[1.99997pt] r\leq z_{i}+\gamma+\xi_{i}^{\top}x\ \ \text{for }i=1,\ldots,n,\\[1.99997pt] \|x\|_{*}\leq v,\\ v\geq 0,\ \ r\geq 0,\ \ z_{i}\geq 0\ \ \text{for }i=1,\ldots,n,\end{array}\right\}, (3) where \|\cdot\|_{*} is the dual norm tied to the ground metric. While ( 3 ) is computationally convenient, selecting \delta is delicate: too small jeopardizes feasibility under shift; too large kills performance. Our contribution is a data-driven selection via a shift-aware GS validator that yields feasibility guarantees under serial dependence. 2 Related Work and Contributions Wasserstein DRO delivers tractable, statistically meaningful robustness [ 9 , 6 ] , with applications to finance and CVaR portfolios [ 2 , 9 , 11 ] . However, calibrating the radius can be challenging and often leads to overly conservative solutions. To avoid these overly conservative solutions, Lam and Qian [ 8 ] develop a normalized Gaussian-supremum validator (among other validators) for data-driven optimization with uncertain constraints. Instead of fixing a very conservative reformulation (large safety margins), their approach generates a solution path of candidate portfolios indexed by a conservativeness parameter (e.g., different values of \rho in the DRO formulation). They then evaluate these candidates on a holdout dataset to find the least conservative solution that still meets the constraint with high confidence ( 1-\beta ). The GS validator computes a simultaneous confidence bound across the entire solution path by estimating the quantile of the supremum of Gaussian deviations. In machine learning, covariate shift refers to the scenario where the input distribution changes between training and testing phases, while the conditional relationship (e.g., returns given factors) remains stable. A rich body of work addresses this via importance weighting techniques: direct ratio estimation (KLIEP, uLSIF) [ 13 , 5 ] , kernel mean matching (KMM) [ 4 ] , and classifier-based ratios [ 1 ] ; these ideas enable targeting a shifted law in validation. Furthermore, standard statistical inference assumes independent data, but financial time series exhibit autocorrelation, volatility clustering, and other dependencies. To overcome this, time-series inference under dependence uses block bootstraps [ 7 , 10 ] and multiplier/wild bootstraps for dependent data [ 12 , 3 ] . Our validation builds on the Gaussian-supremum validator of Lam & Qian [ 8 ] ; we extend it to non-i.i.d. financial data with shift-aware weighting and BMB calibration. Contributions. (i) A shift-aware GS validator : we reweight validation data by estimated density ratios to evaluate CVaR under a learned center Q , then perform Wasserstein DRO around Q with a smaller radius. (ii) A dependence-robust calibration : we replace i.i.d. Gaussian calibration by a block multiplier bootstrap for the GS band, yielding correct joint coverage under serial dependence. (iii) Finite-sample feasibility : we prove that any selected portfolio meets the CVaR limit under Q with probability \geq 1-\beta ; proofs follow GS arguments with extensions for weighting and dependence. 3 Methodology and Main Results We adopt a two-phase procedure. Phase I generates a candidate path \{x^{*}(\delta_{j})\}_{j=1}^{p} by solving ( 3 ) over a grid of radii \delta_{1}<\cdots<\delta_{p} . Phase II validates feasibility on a holdout, time-ordered sample \{\xi_{t}\}_{t=1}^{n_{2}} potentially from a shifted regime. The full algorithm is detailed in Appendix A . Shift-aware center via density ratios. Let P denote the holdout law and Q the target law for deployment (current regime). We estimate w(\xi)=q(\xi)/p(\xi) using a probabilistic classifier that distinguishes recent (proxy for Q ) vs. holdout (proxy for P ) samples, convert scores to odds, clip , and normalize: \tilde{w}_{t}=w_{t}/\sum_{i}w_{i} . For any candidate x , we estimate the CVaR functional’s expectation under Q by the weighted mean: \widehat{\Psi}_{Q}(x)\;=\;\min_{t\in\mathbb{R}}\left\{t+\tfrac{1}{\alpha}\sum_{i=1}^{n_{2}}\tilde{w}_{i}\,\max\{-\xi_{i}^{\top}x-t,\,0\}\right\}. Effective sample size is n_{\text{eff}}=1/\sum_{i}\tilde{w}_{i}^{2} ; we abstain if n_{\text{eff}} collapses. Dependence-robust GS calibration. For each candidate x_{j}=x^{*}(\delta_{j}) with empirical minimizer \hat{t}_{j} , define scores \phi_{ij}=\hat{t}_{j}+\frac{1}{\alpha}\max\{-\xi_{i}^{\top}x_{j}-\hat{t}_{j},0\} and weighted means \widehat{H}_{j}=\sum_{i}\tilde{w}_{i}\,\phi_{ij} . We form centered block sums R_{kj}=\sum_{i\in\mathcal{B}_{k}}\tilde{w}_{i}(\phi_{ij}-\widehat{H}_{j}) over contiguous \{B_{k}\}_{k=1}^{K} of length b (with b\to\infty,\,b/n_{2}\to 0 ). Drawing i.i.d. Gaussian multipliers \varepsilon_{k} , the bootstrap max statistic is T^{*}=\max_{1\leq j\leq p}\ \frac{1}{\sqrt{n_{\text{eff}}}}\ \frac{\sum_{k=1}^{K}\varepsilon_{k}\,R_{kj}}{\widehat{\sigma}_{j}},\quad\widehat{\sigma}_{j}^{2}=\sum_{i}\tilde{w}_{i}\,(\phi_{ij}-\widehat{H}_{j})^{2}. Let \widehat{q}_{1-\beta} be the (1-\beta) -quantile of \{T^{*(b)}\}_{b=1}^{M} . We build a simultaneous upper band for the robust constraint: U_{j}\;=\;\underbrace{\widehat{H}_{j}}_{\text{empirical (weighted)}}\;+\;\underbrace{\tfrac{\delta_{j}}{\alpha}\,\|x_{j}\|_{*}}_{\text{DRO margin}}\;+\;\widehat{q}_{1-\beta}\,\frac{\widehat{\sigma}_{j}}{\sqrt{n_{\text{eff}}}}, and declare x_{j} feasible if U_{j}\leq\gamma . We select the least conservative feasible radius \delta^{*}=\min\{\delta_{j}:U_{j}\leq\gamma\}

Read the complete text