Ensembles That Disagree Where the Price Optimizer Looks

A retailer that sets price and stock level together has to predict demand at prices it has rarely or never charged. The optimizer that picks the price actively searches for the spots where the demand model is most optimistic. Those spots are often where the model is simply wrong. This proposal studies model ensembling (training several different demand models and combining them) as a way to catch that failure, and argues that the usual recipe fails here. Copies of one network trained from different random seeds, or the trees inside a gradient-boosted model, tend to agree with each other off the logged price range, because they extrapolate in the same way. Their agreement then looks like confidence exactly where there should be doubt. We propose to build demand ensembles whose members extrapolate differently in price, to measure their disagreement at the decisions the optimizer actually proposes, and to separate the models that choose a decision from the models that evaluate it, so the reported profit is not inflated by the search.

The stakes are concrete. Prior work has already shown one hidden optimism bias in joint pricing and inventory: plugging mean demand into a concave profit function overstates profit by Jensen's inequality. Deep Generative Demand Learning for Newsvendor and Pricing fixes this by sampling full demand distributions from a single conditional generator. A second optimism bias is left untouched: taking the maximum over prices of any single model's profit estimate rewards that model's errors. This "optimizer's curse" is an epistemic problem, meaning it comes from the model not knowing, rather than from noise in demand. One model cannot see its own epistemic errors, so the fix has to come from comparing several models.

Core Claim

In joint pricing and inventory, an ensemble's value depends on whether its members disagree about the price response, meaning how demand moves with price, at prices outside the logged support. Their disagreement on in-sample demand is the wrong thing to look at.

We make three testable claims. First, disagreement within a tree ensemble is nearly zero off the logged price range, because trees predict a constant beyond their last split. Seed-diverse deep ensembles also show low disagreement there, for a different reason: they collapse toward a common solution once they fit the training data exactly. Either kind of ensemble therefore gives little warning about bad price extrapolation. Second, an ensemble built for extrapolation diversity will have disagreement that tracks decision regret. Its members would be a constant-elasticity parametric model, a network constrained so demand falls as price rises, a conditional generator of the kind in the generative newsvendor work, and a tree model. Third, if we choose the price and order quantity with one half of the ensemble and score them with the other half, we remove most of the gap between estimated and realized profit. Adding a penalty on disagreement to that choice lowers regret relative to the single-generator method.

The contrast with prior work is specific. When are ensembles really effective? shows that the benefit of an ensemble grows linearly with its disagreement-error ratio: expected disagreement between members divided by expected member error. That paper tests this with correlation R ≥ 0.96 on classification benchmarks. It also shows that interpolating deep networks lose disagreement, while random forests keep it. We take this diagnostic from that paper and move it from prediction error to decision regret. We predict that the forests' advantage reverses under price extrapolation. Dual-Agent Deep Reinforcement Learning for Dynamic Pricing and Replenishment uses a LightGBM tree ensemble only as a fixed point-prediction demand simulator for its pricing agents, and never looks at disagreement within it. This is the exact setting where our first claim predicts silent overconfidence.

Research Program

The main object of study is the single-period newsvendor-with-pricing problem. A seller sees context $x$, picks price $p$ and order quantity $y$, and earns concave profit $\Pi(p,y,D)$ on random demand $D$. We start here because it isolates the pricing optimizer's search from multi-period dynamics, and because it has the benchmarks and baselines from the generative newsvendor paper. Each ensemble member $m$ is a model that can produce demand samples at any $(x,p)$. From those samples it gives a profit estimate $\hat\Pi_m(p,y)$. Inside each member, we set the inventory to the critical quantile of its own samples, following the generative newsvendor method.

We split the $M$ members into a selection set $S$ and an evaluation set $E$. We choose

$$(p^\*, y^\*) = \arg\max_{p,y}\; \frac{1}{|S|}\sum_{m\in S}\hat\Pi_m(p,y) \;-\; \lambda\,\mathrm{sd}_{m\in S}\,\hat\Pi_m(p,y),$$

Here sd is the standard deviation of profit estimates across the selection members, and $\lambda \ge 0$ sets how much disagreement is penalized. We then report the evaluation set's average $\bar\Pi_E(p^\*,y^\*)$ as the profit estimate. The quantity we track is the optimism gap: estimated profit minus realized profit on held-out demand. With $\lambda=0$ and $S=E$, this gap reproduces the optimizer's curse. Cross-fitting should shrink the gap, because $E$'s errors are independent of the search that $S$ drove.

The analysis builds a decision-level analogue of the disagreement-error ratio. It is the ensemble's profit disagreement at the chosen decision divided by the mean regret of individual members. We ask whether this quantity predicts how much the ensemble reduces regret across data sets, the way the original ratio predicts accuracy gains.

The construction question is how to create diversity in extrapolation. We compare three families: seed-only deep ensembles; tree ensembles; and heterogeneous ensembles whose members differ in the functional form they impose on the price response. The heterogeneous family borrows from Adaptive expert fusion model for online wind power prediction, which mixes tree and sequence experts and weights them by local error on similar recent points. We adapt that local weighting by defining neighbors in price as well as context. Near a candidate price, members that were accurate get weight. Far from all logged prices, weights fall back to uniform, so disagreement stays visible.

The multi-period extension reuses the setting of Offline Dynamic Inventory and Pricing Strategy: Addressing Censored and Dependent Demand. That paper learns pricing and replenishment policies from offline logs where stockouts hide true demand. It handles this with Kaplan-Meier imputation, and its pessimistic variant (PC-FQI) adds a penalty for uncertain actions. We replace that penalty with disagreement across price-response-diverse Q-function members. Nothing else in the method changes, so the comparison isolates the source of pessimism.

Baselines

The naive reference point has two parts, because it must cover both biases. The first is sample average approximation, which picks the best price and order from the empirical distribution of logged demand. The second is a single mean-demand regressor with inventory set at its point prediction. The generative newsvendor paper reports sample average approximation at a profit gap of 32.02 from the oracle on its hardest synthetic process (DGP 'd'), so any serious method must beat this easily.

The strongest existing competitor is the single conditional generator from Deep Generative Demand Learning for Newsvendor and Pricing. It already solves the mean-demand bias, and it reports a profit gap from the oracle of 2.64 on DGP 'd'. That compares with 14.55 for its kernel baseline and 32.02 for sample average approximation. In its continuous-price joint simulation it reports profit of 103.54 against 78.95 for the kernel method. We will rerun it with its reported architecture and tune it on the same validation splits. Its hyperparameter search budget will be set to match the total compute of our whole ensemble, so a larger single generator gets the same resources.

Beating it means two things. On the paper's original DGPs, where test prices lie inside the logged range, we should roughly match it, since our claim predicts little gain there. On logs where the training prices are confined to a band and the oracle optimum lies outside it, we should achieve a lower profit gap from the oracle and a smaller optimism gap. Both differences must hold across seeds, measured by paired confidence intervals over repeated draws of the data.

In the multi-period setting the strongest competitor is PC-FQI, run with its own penalty on the same offline logs and with the same number of fitted-iteration steps. The file reports convergence within 5 to 50 training episodes but no regret number. We therefore report our comparison as relative regret against the oracle, on the same simulator.

The key ablations isolate each ingredient. First, the full heterogeneous ensemble against a seed-only deep ensemble of equal size, which tests whether diversity in extrapolation matters. Second, against a LightGBM tree ensemble scored by the spread of its trees, which is the uncertainty FSDA-style simulators could have used. Third, with and without cross-fitting, holding $\lambda$ fixed. Fourth, with $\lambda$ set to zero, which separates the gain from honest evaluation from the gain from penalizing disagreement.

Why not just…?

Why not just restrict prices to the range seen in the logs? This is the most serious threat, and we treat it as a baseline, not a straw man. If the true optimum lies inside the logged band, clipping prices to that band is simple and nearly free. Our method should then do no better than clipping, and we will say so. Clipping cannot help when the optimum lies moderately outside the band. That happens when a retailer has only ever discounted lightly, or when the business wants to test a new price level. We design the price-band experiments so the distance from the band to the true optimum is a controlled variable. The prediction is that clipping ties or wins near the band and loses as the distance grows. The disagreement penalty should move out only as far as the members still agree. If clipping wins everywhere, the proposal's practical claim fails, and the result would still be worth publishing.

Why not just use kernel or Gaussian-process confidence bounds, as in contextual bandits? Nonparametric Contextual Pricing and Inventory Learning under Censored Demand builds principled kernel upper confidence bounds with minimax-optimal regret for joint pricing and inventory. It is an online method. It works with low-dimensional context: its experiments use $d \in \{1,2\}$, and its regret exponent worsens as dimension grows. In the offline generative newsvendor comparison, kernel demand estimates trailed the generator, 14.55 against 2.64 on DGP 'd'. We will include a Gaussian-process demand model with its posterior spread as the uncertainty signal. If it matches our ensemble on the low-dimensional synthetic processes, we will report that. Our argument for ensembles rests on the higher-dimensional Kaggle and marketplace-style features.

Why not just ensemble several seeds of the best generator? This is the cheapest ensemble, and it tests the heart of our claim. The interpolation result in When are ensembles really effective? predicts that seed copies will agree too much once they fit the training data. The seed-only ablation measures this directly. If seed diversity yields decision-level disagreement that predicts regret as well as heterogeneous members do, the extrapolation-diversity claim is wrong, and the simpler recipe should be preferred.

Why not just use distributional reinforcement learning? A Simulation Environment and Reinforcement Learning Method for Waste Reduction models the distribution of rewards and reaches 194.1% human-normalized profit against 146.5% for a standard deep Q-network under high volatility. That captures noise in demand, not the model's ignorance about unseen prices, and its method does not set prices. This separation is exactly what our work targets.

Proposed Experiments

The first experiment modifies the synthetic data processes 'a' through 'd' from the generative newsvendor paper. Logged prices are confined to a band whose distance from the oracle optimal price we vary. We measure profit gap from the oracle, the optimism gap, stockout rate, and leftover units. The decisive check is a plot of decision-level disagreement against how much each ensemble reduces regret, across processes, band widths, and ensemble families. The claim holds if the heterogeneous ensemble shows a strong positive relation, while tree and seed ensembles show near-zero disagreement off the band despite large regret.

The second experiment uses the Kaggle Food Demand Forecasting data, which records checkout price, base price, and promotions across 77 fulfillment centers. Following the generative newsvendor study's purchase-cost and salvage-value settings, we hold out weeks whose checkout prices fall outside the training range. We evaluate realized profit under several cost settings. The file gives only qualitative results for the generator on this data set: no reported absolute profit number was found on Lacuna for the Kaggle food-demand setting. We therefore rerun that baseline ourselves.

The third experiment ports the ensemble penalty into the offline censored multi-period simulator from the PC-FQI paper and compares regret against PC-FQI. It also checks that stockouts, which hide true demand, do not break the diagnostic: censored periods should raise disagreement rather than hide it.

A final stress test asks whether an FSDA-style pricing agent trained against a single LightGBM demand simulator exploits that simulator's flat extrapolation. We compare that agent with one trained against our ensemble with the penalty, both evaluated on a different, held-out demand model. This tests the silent-overconfidence prediction directly on the architecture where it would matter most.