The EVT hedge cuts losses on the worst 0.1% of payoff paths, while ordinary paths barely change. Pérez Octavio Mendoza and Godin make that distinction in their abstract: similar behavior under normal conditions, lower hedging losses in the extreme-scenario evaluation. It is a fair distinction. They have found a cleaner way to train a neural hedge against extreme losses. Whether the reported stress-sample improvement comes from the method or from one trained network remains open.
What does autodiff buy?
At a confidence level of 0.999, empirical CVaR takes the average of the worst 0.1% of simulated losses. A training sample of 50,000 paths supplies just 50 numbers; 400,000 paths supply 400. Those few observations carry every gradient step.
Extreme value theory (EVT) lets the estimator use more of the sample. Exceedances above a sufficiently high threshold are approximately generalized Pareto. Fit that tail over a larger sample slice, then calculate CVaR from a closed-form formula. The authors choose among 20 candidate threshold percentiles between 0.80 and 0.99. Each gets an Anderson-Darling fit test. ForwardStop, a sequential stopping rule applied to the test p-values, then selects the threshold. The method of moments estimates shape and scale; training reverts to empirical CVaR if no threshold survives.
Davar et al. (2025) introduced this approach as POTPG, using forward finite differences for the gradients. Each parameter requires an extra objective evaluation. With two parameters, the cost is small. A network with four hidden layers of 56 neurons makes it restrictive. The new method fixes the threshold and uses automatic differentiation to pass through the losses, moments, fit and CVaR formula in one pass. The authors' proposition establishes that this gives the exact local gradient of the finite-sample objective. They also acknowledge that it does not establish consistency for the true CVaR gradient. Speed motivates the paper, though we found no runtime comparison against the finite-difference version.
For the controlled test, losses follow an exact generalized Pareto distribution with shape 0.4. Two parameters determine the scale, and their optimum is (2, -1). The EVT procedure starts at (4, -3), uses 2,000 draws per iteration and reaches about (1.99, -0.99) after 100 iterations. Since the data follow the estimator's assumed model exactly, this is a check that the machinery works. Only a figure gives the empirical-CVaR endpoint.
The 17.49-to-10.03 drop comes from extreme-payoff paths
The hedging experiment uses a single Merton jump-diffusion, with S0 of 100 and a 140-strike European call rebalanced daily over 63 rebalances. The annual jump arrival rate is 0.1, with mean log-jump +0.4. This far-out-of-the-money call rarely pays. When it does, the payoff can be large. Both hedgers train at alpha 0.999.
Across 100,000 general test paths, the policies are much the same. Both have mean hedging error of -1.31. Empirical CVaR has a 95th percentile of 6.23 versus EVT's 6.11, and a maximum of 50.09 versus 48.80. EVT does slightly worse at the median: -0.40 versus -0.63.
The headline improvement appears in a separate sample of 3,000 paths whose unhedged payoff exceeds its own 0.999 quantile. Defining the event through the market avoids conditioning in either hedger's favour. That choice is sound. On these paths, EVT reduces mean error from 17.49 to 10.03 and the median from 16.52 to 9.13. The 95th percentile falls from 49.22 to 35.46; the maximum falls from 114.27 to 84.36.
The authors acknowledge that "these results should not be interpreted as evidence of an equivalent reduction in unconditional risk under the original market distribution." They defend the method on the grounds that EVT "uses the exceedance distribution to model the tail in regimes where empirical CVaR relies on relatively few observations beyond the target quantile."
A puzzle remains.
Roughly 100 of the 100,000 general paths should fall within the extreme-payoff event. That matches the count in the 0.1% loss tail both policies were trained to reduce. The conditional mean of 17.49 is well above the general 95th percentile of 6.23, suggesting these paths probably dominate the loss tail. Yet the paper does not report test-set CVaR at 0.999, the training objective shared by both hedgers. We therefore cannot see whether the conditional improvement carries through to that objective. The closest evidence is validation CVaR over 50 epochs and 4 runs: the means converge to similar levels after about 20 epochs.
One market, one strike, unclear seed count
Retraining uses 50,000, 100,000, 200,000 and 400,000 paths. At every sample size, EVT has lower conditional errors, shown in boxplots without numerical values. The authors write that the experiment "does not support the stronger claim that increasing the number of training trajectories systematically improves either policy over this range." That matters for the paper's motivation. Empirical CVaR gets 50 tail observations at the smaller sample size and 400 at the larger one, without clear improvement. At these sizes, optimization noise appears to outweigh estimation noise, the problem EVT is intended to address.
The training curves cover 4 runs. The hedging-error figures leave the number of trained networks behind each column unstated.
The shape constraint also needs attention. As the authors disclose, the moment estimator requires shape below 1/2, and their estimator clips fitted shape to [0, 1/2). Threshold selection nevertheless accepts fits with shape as high as 0.99. Training can therefore use a capped tail after threshold selection accepts a heavier fitted one. The +0.4 log-jumps could put this market in that region. Without reported fitted shapes, we cannot tell how frequently the cap binds. We also found no transaction costs in the hedging setup. Nor does the paper report whether position sizes differ between the policies.
Our SPY build replaces the simulated Merton market
We are building the comparison with SPY options and SPY shares, using end-of-day option prices and daily hedging. The paper contains simulated experiments and no historical backtest. Our run will therefore supply new evidence; it replicates nothing in the paper. Daily hedging also says nothing about finer rebalancing frequencies. Because the method targets the tail of hedging losses without exploiting a SPY-specific premium, it can carry over in principle. The paper's Merton results cannot carry over with it.
We keep contracts and execution costs identical across the comparison, along with the leverage limit, network and training budget. Only the objective changes. We will report full-sample hedging error and CVaR separately from stress events defined in advance using the unhedged payoff. Lower full-sample CVaR after costs across several seeds would be evidence that the EVT objective helps on SPY options. It would tell us nothing about the paper's Merton networks.