Habibi's headline estimator gives a risk desk a familiar recursion for estimating VaR, with no empirical evidence that it improves the job. The update is a stochastic-gradient quantile recursion in the Robbins-Monro style. Remark 2 acknowledges that Hong and Liu (2021) and Bardou et al. (2020) already supply the same recursion. The remaining question is practical: do the pages that follow, numbered 220 to 227 in the running header, change how you would estimate a 95% one-day tail on a real book?
Following the exceedances
Begin with a guess for the loss quantile at confidence level α, then draw a simulated loss. A loss below the guess nudges it down. A loss above it pushes it up by a much larger amount. Those relative adjustments make the guess settle where exceedances occur at a rate of 1-α. With α = 0.95, drift stops when losses exceed the guess exactly 5% of the time. This fixed point gives VaR. A positive step-size sequence makes the adjustments shrink, though the conditions imposed on that sequence are missing from the text available to us.
The same draws also supply CVaR. Habibi uses the Knight identity to express CVaR through VaR and the loss survival function, then differentiates with respect to the VaR level. The resulting first-order differential equation ties the movement in CVaR to the hazard rate at VaR. An exponential loss has a constant hazard, leaving a fixed gap between CVaR and VaR. The paper also derives closed forms for Pareto and for a Weibull with shape parameter 2. The latter expresses CVaR using the CDF and density of a zero-mean normal.
Several brief extensions surround this core. The paper compares a CLT for the recursion with the Bahadur representation of the empirical quantile. Proposition 2 supplies a continuous-time limit on Brownian motion over [0,1]; Proposition 3 introduces an importance-sampling likelihood ratio. Randomized step sizes appear as bootstrap or uniform weights, and a Newton-Raphson interpretation addresses sensitivity. A translated gamma, matched on 3 moments (mean, standard deviation, skew), then yields VaR from chi-squared quantiles.
The body reports no empirical result of its own.
We found no dataset or sample period, nor any simulation table or runtime figure. What the paper supplies is a numerical estimator for a quantile that could otherwise be read from a sorted vector.
Where was accuracy measured?
The abstract promises to "demonstrate their advantages over conventional methods through extensive empirical validation". It also says the methods "enhance computational efficiency" and "improve the accuracy of VaR and CVaR estimates, particularly in tail regions". The body supplies 3 distributional examples instead, each deriving a closed form for a single parametric loss. Such a formula establishes the correct answer. It cannot establish whether the recursion gets there faster, or closer, than the empirical quantile.
A convergence-rate ratio and a variance ratio are stated for the recursion relative to the Bahadur quantile. Neither expression survived extraction; the paper gives no numerical value. The Bahadur variance depends on the density at VaR. That density is thin in precisely the tails singled out in the abstract.
Three other restrictions affect desk use. The derivation relies on "L is a positive variable", whereas daily P&L losses are signed. The translated gamma requires skew to be "strictly positive". For Proposition 1, Habibi says "the proof is omitted". Sourcing slips further weaken confidence: Sun and Hong (2019) appears as the 2011 Winter Simulation Conference, while the header lists pages 1-15 against the running header's 220-227.
The Langevin target remains unrecoverable
The gradient recursion moves an estimate toward a root. The Langevin proposal adds a standard normal innovation, making the iterate sample a density rather than settle at a point. Adding that noise to a VaR point estimator calls for a reason, starting with a target density and an ergodicity condition. We cannot recover the target density from the available text. We found a discretized update, but no convergence statement or comparison against the plain recursion it would need to beat. Barbu and Zhu (2025), whose work Habibi says he extends, place these samplers within Bayesian learning. Habibi does not identify the posterior that would apply to a single portfolio's VaR.
Our SPY reconstruction
We could not implement the author's method. Many equations are absent from the extracted text, and the Langevin target and settings lack enough detail. We therefore built a reconstructed stochastic-gradient estimator alongside a historical-simulation comparator. All our figures refer only to those constructions. The body provides no benchmark against which to assess the paper's claims of accuracy and efficiency.
We made the construction ourselves. At each SPY close, it uses the trailing 750 signed daily losses and draws 20,000 bootstrap scenarios. The recursion runs for 10,000 steps, using step 0.1/(k+100)^0.6. We average iterates 5,001 to 10,000, then estimate CVaR from the other 10,000 draws. Exposure follows min(1, 0.02/CVaR), targeting 2% estimated daily tail risk without leverage. The book trades at the next open and pays $0.004 a share in commissions; cash earns zero. The window covers 2020-01-01 to 2024-07-01. We use no Langevin noise.
Our gradient-sized SPY book returned 39.44% in total over 2020-01-01 to 2024-07-01. Sharpe was 0.64 and Sortino was 0.75. Volatility came to 13.70%, with a max drawdown of -29.15% and Calmar of 0.26. These results test our substitute construction only. For a book intended to cap tail risk at 2% a day, a 0.64 Sharpe alongside a drawdown of almost 30% is weak.
The paper reports no performance of its own for comparison. Our exposure cap is 100%, and idle cash earns zero, so this book can only hold less exposure than SPY. Its 13.70% volatility should be read before its 39.44% return. We report only the gradient book's figures here.
A single table would change our view of the paper: tail RMSE against draw count, comparing the gradient and Langevin estimators with the plain empirical quantile at equal budgets. Habibi's Pareto case already provides the closed-form answer needed for that comparison.
Our backtest stops at 2024-07-01, and everything after that date is deliberately left untouched so the same strategy can be checked out of sample later.