The certificate covers 4.0 to 8.5 on a stock last quoted at 6.05, yet exactly zero is earned across 3.0 of those 4.5 dollars. Halidias gets there with a written piecewise-affine claim, a stated price range and payoff inequalities checked at four points. Bid and ask remain in the formulation, and the program stays linear. Four linear inequalities and one tail-slope inequality resolve the deployed example. Its economics rest entirely on the width of the stated range.
The main object is a calibrated predictive region, a set of horizon prices intended to contain the future price with probability roughly 1 minus alpha. In the illustration, the investor is short a claim, a call struck at 5.3. The sign parameter is eta = -1, making the constraint a superhedge. A static option portfolio must dominate the claim payoff throughout the region. Losses outside it are capped at a selected level D. The premium is earned when the terminal price falls inside the region, while the floor limits the outcome elsewhere.
The forecast layer uses a one-hidden-layer logistic network (two units, regularization 1e-2, seed 123) fitted to the 23-day log return. Absolute residuals from a chronologically later calibration block produce a finite-sample corrected order statistic, which serves as the radius. Exponentiating the symmetric return interval produces the price interval. The LP then takes that interval, the traded strikes and the claim's breakpoints, and chooses positions in cash, the underlying, calls and puts. The deployed construction is a MILP: it first maximizes capped boundary exposure, then minimizes setup cost. Separate training, calibration and test blocks have 23-origin embargoes, ensuring that every response required by one block is observed before evaluation of the next begins.
The dataset contains 1128 adjusted daily observations of HTZ from 9 November 2021 to 8 May 2026. Lagged-return, realized-volatility and moving-average-gap features leave 1042 labelled origins; 996 remain after the two embargoes. Across 200 purged test origins, the regions covered 200 of 200 at a nominal 90%. The deployed hedge carries two certificates: Profit >= 0 on [4.0, 8.5] and Profit >= -1 for every x >= 0. These figures describe coverage and certification. The paper reports no P&L, and says so.
Piecewise affinity makes the reduction exact. Both the market payoff and the claim are piecewise affine, as is their net payoff. Its minimum on an interval must occur at an interval endpoint, a traded strike or a claim breakpoint. No grid search is needed. Four nodes determine the deployed hedge: the region boundaries 4.0 and 8.5, the claim breakpoint 5.3 and the traded strike 5.5. The paper shows five representative certificate rows, then relies on piecewise affinity and the finite complete enforcement set. An unbounded adverse region requires one additional tail-slope inequality. Each position is split into a buy-at-ask leg and a sell-at-bid leg, carrying bid and ask through the calculation while preserving linearity net of frictions. For multiple assets, boxes are refined by every strike and breakpoint, with checks at 2^d vertices per box. The paper gives no figure for that cost.
The forecast loses to doing nothing
On the 200 untouched test origins (20 June 2025 to 7 April 2026), the sigmoidal regions covered 200 of 200 at a nominal 90%. Their return radius was 0.662819 and their mean price-interval width was 7.129464. An unconditional zero-return forecast calibrated on the same dates covered 192 of 200, with radius 0.483726 and mean width 5.660217. The conditional model is 26% wider and lands farther from its own target. The paper states the result plainly: "The conditional rule is materially wider. These figures therefore do not establish predictive superiority; they document the behavior of one replaceable forecasting front-end."
The second sentence supplies the defence. Remark 3.2 formalizes it: any two forecasting procedures that produce the same region lead to identical construction problems, so the forecaster is declared replaceable. The architectural claim is fair. Width still decides the economics. The displayed mean certified band is 7.129464, against 5.660217 for the trivial benchmark, while 3.0 of the deployed region's 4.5 dollars earns zero. Calling the front end replaceable transfers the burden to a narrow replacement. The paper displays the wide one.
Width also appears in the paper's headline claim. The abstract reports "the raw price interval [4.0258,8.3385], which is enlarged outward to [4.0,8.5]" and promotes a region-binding writer hedge with a global loss bound. A 4.31-wide region on a 6.05 stock is therefore presented as a result, though economically it functions as a cost. The discussion lists five limitations. First among them is the possibility that the region loses calibration during regime shifts.
What the 1.49 purchases
For deployment, the refit uses 811 early labelled origins ending 5 May 2025 and 208 recent calibration origins from 9 June 2025 to 7 April 2026. The corrected rank is 189. The output gives a last close of 6.0500001907, a point forecast of -0.0432602973 and a radius of 0.3640831743. Its raw 90% region is [4.0257645052, 8.3384819593], extending from 33% below the last close to 38% above it. Aligning outward to the half-dollar option grid produces [4.0, 8.5], a certified interval 4.5 wide. The direction of alignment is correct and explicit: "Any alignment to a traded strike grid must enlarge this raw interval outward, never inward."
The MILP selects 0.20 in cash, one long 5.5 call, two short 8.5 calls and two long 9.0 calls. With long positions bought at ask and short positions sold at bid, setup cost is 0.20 + 1.03 - 0.20 + 0.46 = 1.49. The premium is assumed equal to 1.49, with B = 1. As a result, the displayed profit reflects payoff shape without any pricing edge. It is 0.20 on [4.0, 5.3], declines to zero over [5.3, 5.5], and remains flat at zero on (5.5, 8.5]. Three of the certified interval's 4.5 dollars pay nothing. The global floor is -1.
Just beyond the upper boundary, Profit(8.6) = -0.2.
The quotes explain much of the result. The 9.0 call is quoted at 0.04 bid / 0.23 ask, so purchasing two at ask costs 0.19 more than mid. Selling two 8.5 calls at 0.10 against a 0.15 mid surrenders 0.10. Buying the 5.5 call at 1.03 against 0.94 mid costs another 0.09. The package therefore contains 0.38 of half-spread, while its best certified outcome is 0.20. Because the assumed premium equals setup cost, the claim buyer funds the fills and the writer has no initial outlay. The 0.38 shows what execution at mid would have delivered instead. At mid, the package costs 1.11, improving every certified outcome by 0.38 relative to the actual fill. On this menu, the half-spreads (0.38) exceed the best certified outcome (0.20). Halidias accounts for the spreads correctly; the arithmetic belongs to him.
Position size follows from the loss budget and strike grid. With D = 1, a 0.5-wide 8.5/9.0 vertical permits two spreads.
Where the guarantee ends
The paper describes its evaluation of the final payoff on test-block terminal prices as a historical-state stress diagnostic, "not presented as a rolling hedge backtest or as an empirical verification of Proposition 5.7," since contemporaneous option menus cannot be reconstructed for past origins. Proposition 5.7 is conditional-probability bookkeeping. Coverage passes through to the probability of a nonnegative payoff and carries any miscalibration in the forecasting layer with it. The paper also says that the assumptions supporting the standard split-conformal guarantee are not automatic under exchangeability here. It treats the corrected order statistic as an empirical calibration device. Adaptive and sequential conformal work by Gibbs and Candès, Xu and Xie, Stankevičiūtė and coauthors is cited, though none of it is implemented.
One name, one horizon, one forecast origin, one expiration. HTZ is the relisted Nasdaq security. The series includes nothing spliced from HTZGQ, HTZZ or the pre-bankruptcy common stock, and the last close is 6.05. The raw region is 4.31 wide. Strikes are spaced on a 0.50 grid, making 4.0 and 8.5 the nearest available boundaries.
A rolling implementation on liquid underlyings would change my view. I would want a calibrated 90% region narrow enough for the optimizer to select strikes strictly inside it, leaving positive certified payoffs across most of the band. Realized premium would also need to exceed setup cost by more than the quoted half-spreads. The deterministic certificate stands independently of the forecaster. The forecaster remains the part on trial.