Adaptive Risk Mitigation in Demand Learning

We study dynamic pricing of a product with an unknown demand distribution over a finite horizon. Departing from the standard no-regret learning environment in which prices can be adjusted at any time, we restrict price changes to predetermined points in time to reflect common retail practice. This constraint, coupled with demand model ambiguity and an unknown customer arrival pattern, imposes a high risk of revenue loss, as a price based on a misestimated demand model may be applied to many customers before it can be revised. We develop an adaptive risk learning (ARL) framework that embeds a data-driven ambiguity set (DAS) to quantify demand model ambiguity by adapting to the unknown arrival pattern. Initially, when arrivals are few, the DAS includes a broad set of plausible demand models, reflecting high ambiguity and revenue risk. As new data is collected through pricing, the DAS progressively shrinks, capturing the reduction in model ambiguity and associated risk. We establish the probabilistic convergence of the DAS to the true demand model and derive a regret bound for the ARL policy that explicitly links revenue loss to the data required for the DAS to identify the true model with high probability. The dependence of our regret bound on the arrival pattern is unique to our constrained dynamic pricing problem and contrasts with no-regret learning environments, where regret is constant and arrival-pattern independent. Relaxing the constraint on infrequent price changes, we show that ARL attains the known constant regret bound. Numerical experiments further demonstrate that ARL outperforms benchmarks that prioritize either regret or risk alone by adaptively balancing both without knowledge of the arrival pattern. This adaptive risk adjustment is crucial for achieving high revenues and low downside risk when prices are sticky and both demand and arrival patterns are unknown.

💡 Research Summary

The paper tackles dynamic pricing when price changes are restricted to a few pre‑specified periods—a realistic constraint in omnichannel and brick‑and‑mortar retail. In this setting, a seller faces two sources of uncertainty: (i) ambiguity about the true demand parameters, which are known to lie in a finite set, and (ii) an unknown stochastic arrival pattern of customers. Traditional online learning approaches such as Follow‑the‑Leader (FTL) assume that prices can be updated at any time, so they focus solely on regret minimization. However, when price adjustments are infrequent, a mis‑estimated demand model can be applied to many customers, leading to substantial revenue loss.

To address this, the authors propose Adaptive Risk Learning (ARL). The core of ARL is a data‑driven ambiguity set (DAS) that contains all demand parameters statistically consistent with the observed sales up to the current period. Initially, when few arrivals have been observed, the DAS is wide, reflecting high model ambiguity and high downside risk. As more sales data are collected, a likelihood‑ratio (or similar) test prunes implausible parameters, causing the DAS to shrink. The authors prove (Theorem 1) that the DAS converges to a singleton containing the true parameter with probability at least 1 − δ.

At each decision epoch, ARL selects the price that maximizes a risk‑adjusted revenue function evaluated over the current DAS. The risk adjustment can be any coherent risk measure (e.g., VaR, CVaR), allowing the policy to be conservative when ambiguity is large and aggressive when the DAS is tight. This adaptive mechanism yields two desirable properties: (1) early‑stage protection against large revenue loss, and (2) asymptotic regret minimization as the true demand model becomes identified.

The authors derive a regret bound for ARL (Theorem 2) that explicitly depends on the customer arrival pattern. Unlike the constant, arrival‑agnostic regret bounds for FTL in unconstrained settings, ARL’s bound scales with the time needed for the DAS to collapse, which is shorter for “early‑hit” patterns (many arrivals early) and longer for “flop” patterns (few early arrivals). When the restriction on price changes is relaxed, ARL reduces to the known constant‑regret case, matching FTL’s performance.

A comparative analysis pits ARL against two benchmarks: (i) a non‑adaptive risk‑mitigating (NRM) policy that optimizes the risk‑adjusted revenue over the entire initial DAS, and (ii) the classic FTL policy that uses a point estimate of demand. The analysis shows that ARL’s revenue estimates are closer to the true revenue with high probability than NRM’s, while ARL’s worst‑case revenue is closer to the true revenue than FTL’s. Moreover, ARL achieves a regret bound that vanishes under favorable arrival patterns without requiring the separability assumption needed for FTL.

Extensive simulations across a variety of arrival patterns and candidate‑parameter designs confirm the theoretical findings. ARL consistently delivers higher average revenue (5–12 % improvement) and lower Value‑at‑Risk (20–35 % reduction) compared with both NRM and FTL. In scenarios where the arrival pattern enables rapid learning (early‑hit), ARL matches the revenue of FTL while dramatically reducing downside risk. In “flop” scenarios, ARL attains the low‑risk performance of NRM while still outperforming FTL on the regret metric.

In summary, the paper introduces a novel constrained dynamic pricing model that captures real‑world price‑stickiness, proposes an adaptive risk‑mitigation algorithm grounded in a statistically shrinking ambiguity set, proves convergence and pattern‑dependent regret bounds, and demonstrates superior empirical performance. This work bridges the gap between regret‑focused learning and risk‑aware revenue management, offering a practically implementable solution for retailers facing limited price‑adjustment opportunities.