A live book needs evidence that stationary ambiguity improves hedging on real markets, and this paper does not yet supply enough. Its diagnosis is sharp. Its fix requires one parameter, alpha. The supporting backtest is frictionless, however, and reports log risk ratios in figures without error bars we could find. The benchmark set ends with two Black-Scholes deltas.
The authors acknowledge the problem early. Section 5 begins with a concession: "because stationary ambiguity is a general principle for simulator design, it cannot be validated directly on real data." Their proposed research design follows: "Still, we can study the effect of enforcing stationary ambiguity on hedging performance by training policies in simulators with and without stationary ambiguity and backtesting them on historical market data." Fair enough. Yet the abstract promises that policies trained under stationary ambiguity deliver "strong performance on real market data". The sole real-data evidence is a figure showing log risk ratios, without absolute levels.
The hedge learns a regime and trusts it
The mechanism drives the whole argument. Deep hedging trains a neural policy on simulated paths. A standard treatment of parameter uncertainty draws a parameter when each path begins and withholds it from the policy. Mueller, Akkari, Wood and Gonon call this the static latent model (SLM). Volatility, or spot-vol correlation, is drawn once per trajectory from a prior nu.
Returns then reveal the draw to the policy. Under static randomization, the paper proves that expected ambiguity about any square-integrable property of the latent parameter is non-increasing. For Gaussian drift, posterior variance equals tau^2 sigma^2 / (sigma^2 + (t-1) tau^2), which decays at rate 1/t. The policy has effectively settled on a regime before hedging begins and proceeds as if the regime will persist. Linear probes of the LSTM hidden state support that interpretation. Given a warm-up of 64 observations, probe-implied ambiguity about log volatility falls almost monotonically in t for the SLM policy.
A single refresh parameter
The refresh latent model (RLM) redraws the latent parameter from nu at every step with probability alpha. Every main experiment, whether simulated or based on real data, sets alpha to 0.01. The simulator is required to induce a stationary filter process. This condition holds when the joint latent/observation process is stationary Markov, under their Proposition 2. The proposition applies on the bi-infinite time axis; a finite warm-up provides the practical approximation. For the discretized Heston map, the authors verify a uniform contraction condition after Stenflo.
Between refreshes, the base simulator generates each segment exactly. The change therefore amounts to more than adding noise. At T=64, (1-alpha)^T is about 0.53, leaving roughly half the training paths without a jump. Ambiguity inferred from probes of the RLM policy falls initially, then settles above zero.
LSTM policies are trained in the Buehler et al. setup by minimizing an exponential spectral risk measure with gamma=4. All three problems use T=64 and dt=1/250. Two concern ATM straddles, one with uncertain GBM volatility and the other under Heston with uncertain rho. The third prices a worst-of call under an uncertain 2-d covariance. For the Heston problem, kappa=8.0, vbar=0.0625 and xi=1.0 remain known.
An analytic exponential-utility investment problem accompanies these experiments and yields the paper's most useful secondary result. Minimax regret over a static drift interval is a familiar robustness prescription. The paper proves that it converges to specialization in the latent drift and then fails the same regime-shift test. Plain minimax value performs no better because the policy collapses to zero investment whenever the drift interval contains zero.
The advantage appears on the stress grid
Deployability turns on one result. With warm-up H=T and evaluation inside the randomized simulators, the policies are effectively indistinguishable. On RLM paths in BS-VOL, the RLM-trained policy records 10.15 (x100 spectral risk), against 10.19 for the SLM policy. SLM paths favor the SLM policy by 10.31 to 10.33. Aggregate gaps across all three problems range from -0.02 to 0.06.
The separation comes from hand-constructed regime pairs. At the pair producing the largest gap, the policies score 14.94 versus 16.42 in BS-VOL, 16.98 versus 18.47 in HESTON-CORR and 3.79 versus 4.77 in BS-COV. The authors put the point well: aggregate simulator evaluations can miss poor robustness properties of a policy.
They also report the price of that protection. At the regime pair most favorable to the SLM, it beats the RLM by 0.51 in BS-VOL (5.42 against 5.93), by 0.35 in BS-COV (0.81 against 1.16) and by 0.05 in HESTON-CORR. With no history in the initial-robustness test, the plug-in Black-Scholes policy slightly outperforms both randomized policies at evaluation volatilities of roughly 0.12 to 0.17, close to the prior mean. Refresh randomization improves tail behaviour following large shifts and gives up performance in calm regimes that have been identified correctly. The trade depends on how often latent parameters move, something the simulator cannot determine.
Alpha is chosen by judgement. Sensitivity is examined only for BS-VOL at T=64, where semi-deviation remains within a 1.44 to 1.48 x 10^-2 band. The real-data exercise retains alpha=0.01 while doubling the horizon to T=128. With T=128, the same alpha leaves about 28% of paths untouched, using the arithmetic behind the paper's 0.53 figure at T=64.
We have previously examined a learned model whose response to an engineered volatility shock fell far short of naive scaling (/articles/dspm-puts-a-volatility-clock-inside-diffusion-noise).
Ten years of S&P 100 windows
The backtest estimates nu by maximum likelihood from 128-day return windows running from January 2006 through December 2015 (alpha_IG=1.63, beta_IG=0.07). It then hedges ten payoffs on overlapping 128-trading-day windows from January 2016 through December 2025. A 32-day warm-up precedes hedging, and losses are pooled across stocks.
Performance appears as log(R_BS-HIST/R). The RLM policy delivers the lowest pooled risk for the six non-vanilla payoffs. It also leads overall within the top 10% of windows ranked by max/min 21-day rolling realized variance. Across those windows, the BS-EWMA delta with lambda=0.94 beats the SLM policy on 5 of 10 payoffs. This is the paper's clearest evidence that vanishing ambiguity creates measurable hedging risk precisely when volatility moves most.
Four constraints limit that conclusion. Transaction costs are absent throughout the paper by construction of the loss function. The universe fixes the S&P 100 as of end-December 2015 and applies that membership over 2016-2025, introducing hindsight into the constituents. Maximum-likelihood fitting produces a prior with a heavy upper tail, so the authors cap training volatility at 1.0 to avoid excessive sensitivity to extreme draws. We also found no standard errors, absolute risk levels or per-year results. Because losses are pooled across overlapping windows for highly correlated large-cap names, the effective sample is far smaller than the number of windows.
The authors state the strongest objection themselves. One policy for each method and payoff is used across every stock and date. An SLM retrained continuously with an updated nu would probably adapt better. They leave that version untested because training a policy for every stock and hedge date is computationally prohibitive.
Their response is that retraining leaves the underlying distinction intact. A retrained SLM still chooses its control on the assumption that ambiguity will disappear and the regime will persist. The argument reaches only so far. A footnote says the difference is likely to be large when the simulated market permits trading in future realized volatility, perhaps through options, while the backtest permits trades only in the underlying. The experiment omits the setting where the expected gap is widest. Its chosen setting also gives the untested retrained SLM its best opportunity to narrow that gap.
Theory and experiment separate once more in HESTON-CORR. The policy directly observes the variance process V_t, although V_t is unobservable in practice. The recipe section establishes stationary ambiguity for latent V as well, and Appendix C.3 supplies the argument for the randomized Heston map. The hedging experiment uses observed V.
Our implementation and its losses
Our platform contains end-of-day option data only. We approximate continuous hedging through rebalancing at the daily close and reduce the large-scale training setup to a reproducible implementation. These results describe our implementable version, rather than an exact reproduction.
More fundamentally, our book differs from the paper's object. The paper evaluates spectral risk from hedging losses for a specified payoff on normalized spot. Its comparators are Black-Scholes deltas, with no premium leg and no P&L. We instead formed a tradable portfolio of long near-ATM straddles on the top 50 US names by dollar volume, entered with 190 to 195 calendar days remaining until expiry. Each episode lasts 128 trading days after a 32-day warm-up. At the close, a 2-layer LSTM with hidden size 64 sets the underlying hedge. Training uses the refresh latent model with alpha=0.01 and the same gamma=4 spectral objective. Costs are 10 bps on option premium, 1 bp for every hedge change, and $0.004 per share in commissions subject to a $1 minimum.
From 2016-01-01 to 2025-12-31, the book returned -66.95% in total. Its Sharpe was -0.18, realized volatility reached 36.77%, maximum drawdown was -81.09%, and trading totaled 28,898 trades.
Those figures belong to our run.
The simulation and stress results above come from BS-VOL Table 2: 10.15 against 10.19 on RLM paths, and 14.94 against 16.42 at the worst regime pair. They report relative tail risk for hedging error before costs. Premiums are unpaid and capital does not compound. For real data, the paper's headline remains a log risk ratio figure without any absolute level.
Volatility and drawdown explain most of our outcome. Our construction compounds heavily overlapping, levered long-premium episodes, with up to 10% per episode and gross leverage up to 4.0. Losses consequently accumulate during the same low-realized-volatility stretches. The paper's pooled-risk statistic contains no equivalent compounding. Buying six-month single-name straddles systematically across 2016-2025 carries negative carry, an exposure removed entirely by the paper's metric. Fewer than half of our episodes were profitable, with a win rate of 46.19%, while gross gains equaled 0.89x gross losses.
Costs move in one direction.
We estimated our prior from index-constituent windows, then used it for names including NVDA and TSLA. The learned hedge likely under-hedges the highest-volatility episodes as a result. We also clipped hedge output to plus or minus two shares per straddle and rebalanced only at the daily close. Residual gamma remains, unlike in the paper's unconstrained policy. The available evidence does not fully explain the gap. Our run generated no SLM or BS-HIST comparison curves, so it did not test the paper's claim. It measured a long-vol book and taught us about that book.
What would change the empirical case?
The theory stands independently. Anyone who believes worst-case optimization over a static interval provides continual robustness should read the minimax-regret specialization result. The empirical case would become more persuasive with a tightly defined comparison: rerun the regime-shift test against a continuously retrained SLM on even a handful of names, include transaction costs in the loss, and publish absolute risk levels alongside the log ratios. For now, the BS-VOL 1.48 belongs to a grid selected by the authors. The BS-VOL 0.04 is the gap visible to the simulator itself.