A rule that trades 2.47 times a day instead of 20.40 saved $1.79 per episode in Kumar's BTC test, with a 95% CI of [-2.21, -1.39] and p < 0.0001. That number is as close to mechanical as an empirical result gets. The $8.66 P&L gap printed next to it is a different animal, and most of it is not cost.
The band trades 2.47 times a day
Kumar sets up a fair fight that the deep hedging literature skips. Comparisons are frequently made against frictionless Black-Scholes delta on simulated paths, a benchmark chosen to lose. Here all six strategies face the same 5 basis point round-trip cost, charged as 2.5bp per trade, on real quotes.
The setup: short one BTC option, hedge in spot BTC, rebalance hourly. Three classical hedges. Black-Scholes delta computed from each contract's own quoted mark_iv. Leland's cost-inflated volatility, with k = 5bp and dt of one hour. And the Whalley-Wilmott no-trade band, which holds the position flat while delta stays inside a band of half-width proportional to (kS\u00b2\u0393\u00b2/\u03bb)^(1/3) and rebalances only to the band edge on a breach. Against those, three neural policies. A 32-unit single-layer LSTM trained on CVaR at the 95% level, and the same LSTM with a turnover penalty of beta = 5. Then a feedforward net, two layers of 32 with a four-step lookback, at beta = 100. All three see five features including the contemporaneous BS delta, and all three emit a continuous position through 1.5\u00b7tanh.
The data is the constraint that shapes everything. Tardis.dev gives away Deribit option chains for the first day of each month; continuous access was quoted at roughly $650 a month, which Kumar says he could not fund. So the sample is 61 separate 24-hour windows from January 2020 to December 2024. Raw, that is 892,912 hourly rows across 22,866 instruments. After filters, 490,229 survive: a two-sided quote was required, which dropped 27.0% of rows, moneyness was held to 0.5 to 2.0, and relative spread was capped at 50% of mid. Train is December 2019 to December 2022, 8,922 episodes. Validation is January to August 2023, 3,520 episodes. Test is September 2023 to December 2024: 11,546 episodes drawn from 16 unique trading days.
On that test set the band wins on P&L, tail risk and cost. Mean P&L -30.16 against -38.82 for BS delta and -38.91 for Leland, CVaR(95%) -864.23 against -885.24, mean cost 7.91 against 9.70. The three neural policies come last on mean P&L and CVaR: -47.34, -47.24 and -50.90, all with CVaR worse than -960. Kumar's own summary line, that all three deep hedging configurations underperform every classical benchmark on every metric measured, is slightly stronger than his Table 1, where LSTM v2's cost of 9.43 and turnover of 0.647 beat BS delta's 9.70 and 0.668 and Leland's 9.68 and 0.666. Every strategy loses money because every strategy is short options into the 2024 rally, so only the ranking means anything.
Where the other $6.87 sits
Subtract the cost saving from the P&L gap. Whalley-Wilmott beats BS delta by $8.66 per episode, of which $1.79 is transaction cost. The remaining $6.87 is gross hedging P&L. Kumar reads it the same way, writing that the comparative advantage of Whalley-Wilmott appears tied to the volatile, trending character of that window. And it is exactly the piece that does not survive: the P&L difference carries a 95% CI of [-3.36, 20.74] and p = 0.164. Kumar's own diagnosis is that 11,546 episodes sit on only 16 independent sampled days, which caps the effective sample size for the block bootstrap.
Kumar's validation window makes the point better than I can. On the calm January to August 2023 period the band has the worst mean P&L of the three classical strategies, -5.06 against -4.94 for BS delta and -4.81 for Leland, and the worst CVaR at -358.26 against -353.55. Its cost edge holds, 3.50 against 4.02, on 2.06 trades against 18.08. Net of the 0.52 cost saving, the band's gross hedging performance there was worse by about 0.64. He reports this himself, calls the P&L result regime-dependent, and returns to it in the limitations. Grant him that. The sharper reading is that the band reliably buys you a cost reduction and hands you a trend exposure you did not ask for, and the sign of the second part is set by the window.
One thing to note about the inspection-tuned lambda = 60, which Kumar flags as calibrated by eye with a sensitivity analysis left to future work. The half-width enters as lambda^(-1/3), so an eightfold error in lambda moves the band by a factor of two. The one other value he reports trying, lambda = 0.01, produced a band nearly as wide as delta's entire possible range and made the strategy trade once per episode.
Did the networks fail, or did the design?
All three neural policies traded 20.44 times out of 24.
BS delta traded 20.40. They were handed BS delta as an input feature and asked to emit a continuous bounded position. Kumar's explanation is the absence, in all three configurations, of any structural mechanism analogous to Whalley-Wilmott's explicit trade or no-trade rule. The turnover penalty did what a smooth penalty does: it shrank trades, with turnover falling 0.685 to 0.647 and cost down 0.54, p < 0.0001. It did not change how often the model traded, so P&L was unmoved (+0.11, p = 0.964) and CVaR slightly worse (-966.23 to -976.87). Raising the penalty twentyfold and dropping recurrence changed P&L by -3.66 (p = 0.506). A penalty that is statistically significant on the thing it directly prices and inert on everything else is a pattern we have written up before, in a 0DTE ranker whose abstention gate never bound out of sample.
So the negative result is real but narrow. Continuous-output policies with delta in the feature set, trained on 8,922 episodes of at most 24 steps against a literature that simulates 10^5 to 10^6 paths, do not discover a discrete trade-or-hold rule. Kumar says as much. What the study does not test, and does not claim to, is an architecture that parameterizes the band directly.
Leland is nearly a duplicate of the delta baseline here, at -38.91 versus -38.82 on mean P&L and 20.41 trades versus 20.40. Kumar predicts this in his own literature section: the adjustment goes economically negligible when k is small relative to sigma\u00b7sqrt(dt).
What we can test
We cannot reproduce this. We have no intraday Deribit option chains and no intraday crypto implied-volatility surface, so the hourly BTC experiment is out of reach and none of the dollar figures above transfer. The mechanism does transfer. Hold the hedge until delta leaves a band, then trade back to the edge. Nothing in that is BTC-specific. The implementable version is end-of-day US listed equity options hedged with their underlying stocks, at daily rather than hourly frequency, which is a different estimator and will produce different band widths and trade counts. The equity-option version is running now.
Whether the $6.87 of gross P&L survives outside a rally is the question. Kumar measures it over 11,546 test episodes spanning September 2023 to December 2024, drawn from 16 unique sampled days inside a single trending window. Sample the test set across independently defined trend and volatility regimes instead, and if the edge holds, the band is doing something beyond trading less. The cost side needs no such defense: a strategy trading 2.47 times per episode instead of roughly 20.4 will, in Kumar's words, almost always cost less under any positive transaction cost assumption.