Estimating financial tail risk without fixing the shape of market shocks

What the researchers estimate

Yuhan Ma, Heng Peng and Feipeng Zhang study conditional expectile-based value-at-risk, abbreviated EVaR in their preprint. “Conditional” means the estimate responds to information such as past returns or other predictors, rather than describing risk with one fixed number.

An expectile minimizes a weighted average of squared prediction errors, with different weights assigned to observations above and below the estimate. Low expectile levels emphasize downside outcomes. Unlike a percentile, an expectile depends on the magnitudes of deviations, not just where observations fall in the distribution. A 5% expectile therefore should not be read as a threshold breached on exactly 5% of days.

The authors seek to balance two approaches. Fully parametric methods impose a specific model, potentially including an incorrect distribution for market shocks. Nonparametric methods relax functional assumptions but can become less accurate and harder to scale as the number of predictors grows.

A two-step approach

The proposed method separates a return into a conditional mean, a positive conditional volatility and a standardized innovation, or shock. The mean and volatility follow specified parametric functions, which can accommodate nonlinear dynamics and ARCH- or GARCH-type volatility models. The innovation distribution is left unspecified.

First, the researchers estimate the mean and volatility parameters using quasi-maximum likelihood. They then subtract the fitted mean from each return and divide by its fitted volatility. These standardized residuals represent what remains after filtering out the modeled dynamics.

Second, they calculate a sample expectile from those residuals. To obtain the conditional risk estimate, they multiply that residual expectile by the predicted volatility and add the predicted mean. This is an application of filtered historical simulation: historical shocks supply distributional information after predictable changes in location and scale have been removed.

The distinction matters. The procedure avoids choosing a complete probability distribution for innovations, but it does not eliminate the need to specify the mean and volatility models.

Measuring uncertainty

Under their stated assumptions, the authors establish an estimation-error rate proportional to the inverse square root of sample size, commonly called the parametric convergence rate. They also derive a functional central limit theorem describing uncertainty across a range of expectile levels, rather than only at one level.

These results require conditions including stationarity, restrictions on serial dependence, finite moments and independent, identically distributed innovations independent of the predictors. The process theory concerns expectile levels within an interval bounded away from zero and one.

For practical uncertainty estimates, the preprint proposes a fixed-design residual bootstrap. It repeatedly resamples fitted residuals, constructs artificial returns while retaining the original predictor design, and refits the two-step estimator. Variation across those estimates supplies a bootstrap standard error.

The framework also extends to expectile-based expected shortfall, defined here by averaging lower-level expectiles, and to reconstructing a conditional return distribution from an estimated expectile curve.

What the numerical evidence shows

The simulations examine four return models, including GARCH and asymmetric GJR-GARCH specifications, with normal or standardized Student-t innovations. Sample sizes are 500 and 1,000. The authors report generally lower mean squared errors and coverage probabilities closer to the intended 95% level than two nonparametric competitors.

The comparison with fully parametric estimation reveals a trade-off. When the innovation distribution is correctly specified, the fully parametric estimator generally has smaller mean squared error. Incorrect distributional assumptions can instead produce substantially worse results. Tests fitting a GARCH model to GJR-GARCH-generated data provide evidence of resilience in those particular misspecification scenarios, not a guarantee against arbitrary modeling errors.

Evidence from the S&P 500

The empirical application uses 2,514 daily S&P 500 log returns from January 2016 through December 2025. The first 1,514 observations form the training sample; the remaining 1,000 support out-of-sample evaluation.

Against one nonparametric benchmark, the proposed estimator records lower realized expectile losses at all seven reported levels. At the 0.05 level, the losses are 0.252 versus 0.255. The authors also report smaller, more stable estimated standard errors.

Distributional forecast scores favor the proposed method numerically across all reported criteria. Statistical evidence is narrower: the reported p-values fall below 0.05 for the predictive log-score and the double-tail weighted expectile score, but not for the other four weighted scores.

The results support the authors’ proposed compromise between structure and flexibility within the settings examined. High-dimensional predictors, machine-learning-based first-stage models and extreme conditional expectiles remain directions they identify for future work.