D-GARA: Tail-Risk Generative Adversarial Regression

• D-GARA is a framework for conditional scenario generation that preserves tail risk by ensuring synthetic policy outcomes match real-world risks, specifically via the VaR–ES pair.
• It employs a generative regression approach with adversarially selected policies, optimizing a strictly consistent score to align risk estimations across varying market conditions.
• Empirical studies on S&P 500 data demonstrate that temporal architectures like Encoder-LSTM achieve near-nominal risk calibration, underscoring the benefit of explicit temporal modeling.

D-GARA is the tail-risk instantiation of Generative Adversarial Regression (GAR), a framework for conditional scenario generation in which generated trajectories are judged by the risk of downstream policy outcomes rather than by generic distributional proximity alone. Its defining objective is risk preservation: for a given context, the conditional tail risk of synthetic policy-induced outcomes should match that of real policy-induced outcomes. In the formulation studied in finance, scenarios are multivariate return paths, policies map those paths into trading actions, and the preserved risk object is the jointly elicitable pair $(\mathrm{VaR}_\alpha,\mathrm{ES}_\alpha)$ of policy-induced PnL (Asadi et al., 9 Mar 2026).

1. Conditional risk scenarios and the target of learning

D-GARA is formulated for conditional scenario generation rather than ordinary point prediction. The data are i.i.d. samples

$\mathcal{D}=\{(c_i,y_i)\}_{i=1}^N,$

with context $C\in\mathbb{R}^{d_c}$ and high-dimensional scenario $Y\in\mathbb{R}^{M\times T}$. A conditional generator $G_\theta$ takes latent noise $Z\sim F_Z$ together with a context $c$ and produces a synthetic trajectory

$Y_\theta = G_\theta(Z,c).$

The downstream decision layer is explicit. A policy $\pi:\mathbb{R}^{M\times T}\to \mathcal{U}^T$ maps a scenario into an action sequence, and a deterministic aggregator $A$ maps the pair $\mathcal{D}=\{(c_i,y_i)\}_{i=1}^N,$0 to a scalar outcome,

$\mathcal{D}=\{(c_i,y_i)\}_{i=1}^N,$1

The real downstream outcome is $\mathcal{D}=\{(c_i,y_i)\}_{i=1}^N,$2, while the synthetic outcome induced by the generator is

$\mathcal{D}=\{(c_i,y_i)\}_{i=1}^N,$3

The central target is not merely

$\mathcal{D}=\{(c_i,y_i)\}_{i=1}^N,$4

but instead

$\mathcal{D}=\{(c_i,y_i)\}_{i=1}^N,$5

This is the paper’s notion of risk preservation under conditional generation (Asadi et al., 9 Mar 2026).

The motivating critique is that conventional conditional generators are trained to match distributions under generic metrics, whereas downstream risk is evaluated only after a policy acts on the trajectory. A generator can therefore appear distributionally close while still being misaligned for risk. In the financial example used in the paper, $\mathcal{D}=\{(c_i,y_i)\}_{i=1}^N,$6 is a multivariate return path, $\mathcal{D}=\{(c_i,y_i)\}_{i=1}^N,$7 summarizes the current market regime, and the downstream scalar is PnL generated by a trading policy. Small discrepancies in distributional regions exploited by the policy can then induce large tail-risk errors.

2. Elicitability and the regression characterization of conditional risk

GAR and D-GARA are built on elicitability. A risk functional $\mathcal{D}=\{(c_i,y_i)\}_{i=1}^N,$8 is elicitable if there exists a scoring function $\mathcal{D}=\{(c_i,y_i)\}_{i=1}^N,$9 such that

$C\in\mathbb{R}^{d_c}$0

When the minimizer is unique, the score is strictly consistent.

Conditionally, for a fixed policy-induced functional $C\in\mathbb{R}^{d_c}$1 and context $C\in\mathbb{R}^{d_c}$2,

$C\in\mathbb{R}^{d_c}$3

This yields the regression problem

$C\in\mathbb{R}^{d_c}$4

whose minimizer satisfies

$C\in\mathbb{R}^{d_c}$5

Accordingly, conditional risk estimation can be cast as learning a regression function under a strictly consistent score (Asadi et al., 9 Mar 2026).

The paper enumerates several elicitable functionals relevant for D-GARA. For a scalar loss $C\in\mathbb{R}^{d_c}$6 with CDF $C\in\mathbb{R}^{d_c}$7, the $C\in\mathbb{R}^{d_c}$8-quantile is

$C\in\mathbb{R}^{d_c}$9

and the $Y\in\mathbb{R}^{M\times T}$0-expectile is the unique $Y\in\mathbb{R}^{M\times T}$1 solving

$Y\in\mathbb{R}^{M\times T}$2

The strictly consistent scores used include the quantile score

$Y\in\mathbb{R}^{M\times T}$3

and the expectile score

$Y\in\mathbb{R}^{M\times T}$4

Expected shortfall alone is not elicitable, but the pair $Y\in\mathbb{R}^{M\times T}$5 is jointly elicitable. The paper therefore uses a Fissler–Ziegel-style joint score. Writing $Y\in\mathbb{R}^{M\times T}$6, a broad class of strictly consistent scores is

$Y\in\mathbb{R}^{M\times T}$7

with $Y\in\mathbb{R}^{M\times T}$8 increasing and $Y\in\mathbb{R}^{M\times T}$9 differentiable, strictly increasing, and strictly convex. In the experiments,

$G_\theta$0

and the indicator is replaced by the smooth surrogate

$G_\theta$1

to enable backpropagation.

3. From risk regression to generative regression and adversarial training

The key conceptual move in GAR is to replace direct pointwise risk prediction by generator-implied risk prediction. For a fixed policy functional $G_\theta$2, the generator induces the conditional risk

$G_\theta$3

Rather than fitting $G_\theta$4 directly, GAR learns $G_\theta$5 so that the generator itself produces the correct conditional risk:

$G_\theta$6

This is the paper’s generative regression objective (Asadi et al., 9 Mar 2026).

If calibration is performed only against a finite benchmark family $G_\theta$7, the objective becomes

$G_\theta$8

The paper explicitly characterizes this construction as useful but brittle, because it guarantees alignment only for the chosen benchmark set.

To remove dependence on a fixed set of policies, GAR introduces an adversarial policy family $G_\theta$9 and trains the generator against the worst-case policy-induced risk mismatch:

$Z\sim F_Z$0

The inner maximization selects the policy under which real and synthetic conditional risks differ most according to the strictly consistent score; the outer minimization adapts the generator to eliminate that discrepancy. The paper’s stated effect is robustness to policy shift across a broad class of admissible decision rules rather than fidelity to a fixed, pre-specified policy set.

4. D-GARA as the VaR–ES specialization of GAR

D-GARA is the practical tail-risk version of GAR focused on the jointly elicitable pair $Z\sim F_Z$1. In this specialization the downstream outcome is typically PnL under a trading policy. For a scenario $Z\sim F_Z$2, the policy outputs time-varying portfolio weights

$Z\sim F_Z$3

with $Z\sim F_Z$4 causal. The policy-induced functional is

$Z\sim F_Z$5

To avoid degenerate leverage, the policy is constrained by fixed gross exposure,

$Z\sim F_Z$6

implemented via

$Z\sim F_Z$7

For a fixed context $Z\sim F_Z$8, the generator produces $Z\sim F_Z$9 synthetic scenarios

$c$0

which are mapped through the policy to synthetic outcomes

$c$1

The generator-implied risk is then estimated by plug-in,

$c$2

which for $c$3 means empirical quantile and empirical tail average. The training objective becomes

$c$4

In this formulation, D-GARA aims to preserve three properties. First, it seeks preservation of conditional tail risk of policy outcomes:

$c$5

Second, it seeks robustness across policies, not merely across benchmark trading strategies but across adversarially selected policies within a class. Third, it seeks context sensitivity, so that the form of risk alignment varies with market conditions (Asadi et al., 9 Mar 2026).

5. Optimization procedure and empirical study on S&P 500 data

Training uses alternating stochastic min–max updates. For each minibatch, the adversary update freezes $c$6, generates synthetic scenarios, computes synthetic and real outcomes under $c$7, evaluates the score, and ascends:

$c$8

The generator update then freezes $c$9, recomputes the loss, and descends:

$Y_\theta = G_\theta(Z,c).$0

Each step uses Monte Carlo samples from the latent $Y_\theta = G_\theta(Z,c).$1 to estimate generator-implied conditional risk (Asadi et al., 9 Mar 2026).

The empirical study uses daily log returns of nine S&P 500 stocks: AAPL, INTC, T, F, BAC, NEE, MU, AMD, and PFE. The date range is 1984-06-01 to 2025-08-20, with an 80/10/10 train/validation/test split. The conditioning window is 5 past daily returns, the generated horizon is 10 future days, the Monte Carlo sample size is 2000, latent noise is $Y_\theta = G_\theta(Z,c).$2, and the risk level is $Y_\theta = G_\theta(Z,c).$3. Three conditional generators are compared: Simple-linear, Encoder-linear, and Encoder-LSTM. Baselines are an Unconditional generator, a DCC-GARCH baseline, and a Direct linear model predicting VaR and ES without scenario generation:

$Y_\theta = G_\theta(Z,c).$4

Evaluation uses the joint VaR–ES strictly consistent score $Y_\theta = G_\theta(Z,c).$5 and the VaR violation rate at $Y_\theta = G_\theta(Z,c).$6, defined as the fraction of times realized PnL falls below estimated $Y_\theta = G_\theta(Z,c).$7. Lower $Y_\theta = G_\theta(Z,c).$8 is better, and the VaR violation rate is ideally close to $Y_\theta = G_\theta(Z,c).$9.

The reported findings are threefold. First, all three conditional generators outperform the unconditional generator, DCC-GARCH, and the direct linear model on the joint VaR–ES score across train, validation, and test. Among them, Encoder-LSTM performs best overall on $\pi:\mathbb{R}^{M\times T}\to \mathcal{U}^T$0, which the paper interprets as evidence that explicit temporal modeling improves risk preservation. Second, at $\pi:\mathbb{R}^{M\times T}\to \mathcal{U}^T$1, the VaR violation rates are: Encoder-LSTM, $\pi:\mathbb{R}^{M\times T}\to \mathcal{U}^T$2; Encoder-linear, $\pi:\mathbb{R}^{M\times T}\to \mathcal{U}^T$3; Simple-linear, $\pi:\mathbb{R}^{M\times T}\to \mathcal{U}^T$4; Direct, $\pi:\mathbb{R}^{M\times T}\to \mathcal{U}^T$5; Unconditional, $\pi:\mathbb{R}^{M\times T}\to \mathcal{U}^T$6; and DCC-GARCH, $\pi:\mathbb{R}^{M\times T}\to \mathcal{U}^T$7. The paper’s interpretation is that DCC-GARCH substantially underestimates tail risk, the other models are conservative, and Encoder-LSTM is closest to nominal calibration. Third, when fixed-policy training is compared with adversarial min–max training, benchmark-strategy performance is similar, but adversarially trained models consistently perform better under worst-case policies; the gain is largest for simpler architectures, while Encoder-LSTM appears naturally more stable.

6. Methodological significance, scope, and terminological ambiguity

The paper presents D-GARA’s methodological contribution as a three-step lift from forecast evaluation to generative modeling: use elicitability to convert conditional risk estimation into regression; replace direct prediction by a generator whose outputs are judged through downstream policy-induced outcomes; and make the policy adversarial so that the learned generator is robust across admissible decision rules (Asadi et al., 9 Mar 2026). This suggests that D-GARA is best understood not as a generic conditional generator, but as a framework for learning risk-aligned conditional scenarios.

Several misconceptions are addressed implicitly by the construction. D-GARA is not aimed at ordinary point prediction, because the object of learning is a conditional law filtered through policy and risk evaluation. It is not equivalent to fitting a generator against benchmark policies alone, because the paper identifies finite-policy calibration as brittle. It is also not a method for expected shortfall alone, because the tail-risk specialization relies on the jointly elicitable $\pi:\mathbb{R}^{M\times T}\to \mathcal{U}^T$8 pair.

There is also a nomenclatural ambiguity. A separate arXiv paper in smart-city parking uses the label DGRA / D-GARA for the Dynamic Gap Reduction Algorithm, a crowdsensing method for roadside parking occupancy prediction under sparse mobile scans (Zheng et al., 2024). That work is unrelated in domain, objective, and mathematical structure. In the GAR literature, by contrast, D-GARA denotes the tail-risk instantiation of Generative Adversarial Regression centered on conditional scenario generation, adversarial policy selection, and preservation of policy-induced $\pi:\mathbb{R}^{M\times T}\to \mathcal{U}^T$9 under context.