Abstract
This paper studies offline dynamic pricing without data coverage assumption, thereby allowing for any price including the optimal one not being observed in the offline data. Previous approaches that rely on the various coverage assumptions such as that the optimal prices are observable, would lead to suboptimal decisions and consequently, reduced profits. We address this challenge by framing the problem to a partial identification framework. Specifically, we establish a partial identification bound for the demand parameter whose associated price is unobserved by leveraging the inherent monotonicity property in the pricing problem. We further incorporate pessimistic and opportunistic strategies within the proposed partial identification framework to derive the estimated policy. Theoretically, we establish rate-optimal finite-sample regret guarantees for both strategies. Empirically, we demonstrate the superior performance of the newly proposed methods via a synthetic environment. This research provides practitioners with valuable insights into offline pricing strategies in the challenging no-coverage setting, ultimately fostering sustainable growth and profitability of the company.
Introduction
In business analytics, a persistent "Catch-22" exists: how can a firm determine the optimal price for a product if that price has never been tested historically? In the context of offline reinforcement learning (RL), this is known as the coverage problem. Most existing offline RL algorithms assume "full coverage"—the premise that historical data is diverse enough to have explored all possible actions, including the optimal ones.
In dynamic pricing, however, managers rarely set prices at random, leading to "narrowed and biased" historical datasets. If the true optimal price was never tested, standard greedy algorithms will ignore it, and even modern pessimistic RL methods—designed to be cautious with uncertain data—typically fail to identify it. This research addresses the "no-coverage" setting, where the optimal policy might be entirely absent from the training data, by leveraging the fundamental economic principle of monotonicity.
Methods
The core innovation of this work is framing the pricing problem as a Partial Identification task. By utilizing the fact that demand generally decreases as prices increase (monotonicity), the researchers can mathematically "sandwich" the potential demand for unobserved prices.
Partial Identification and Monotonicity
If the demand at $10 and $20 is known, the demand at $15 can be bounded even without direct observation. The authors construct confidence intervals for demand parameters at every price point. For unobserved prices, these intervals are derived from the closest observed neighbors.
Strategic Frameworks: Pessimism vs. Opportunism
The paper introduces two distinct strategies within this partial identification framework:
- Refined Pessimism: This strategy assumes the worst-case scenario within the identified bounds. It selects the price that maximizes the lower bound of expected revenue. While robust and suitable for risk-averse scenarios, it tends to favor observed prices where uncertainty is lower.
- Opportunism: Inspired by minimax regret, this strategy chooses the action that minimizes the maximum possible loss (regret). Unlike pessimism, it considers the potential upside of unobserved prices, making it more "adventurous" and capable of identifying unobserved optimal prices.
These strategies are integrated into a backward induction process to handle multi-period pricing and inventory constraints effectively.
Results and Discussion
The authors evaluated their methods using a synthetic environment mimicking a challenging real-world inventory scenario with $T=4$ time periods and very small data ($N=20$ trajectories).
Theoretical Guarantees
The research establishes "rate-optimal" finite-sample regret bounds, proving that as data volume increases, the error shrinks at the best possible mathematical rate ($1/\sqrt{N}$), even in the absence of the coverage assumption. These bounds are shown to be "tight," accounting for the inherent cost of missing optimal price data.
Empirical Performance
In scenarios where the optimal price was intentionally omitted from the offline data, the Opportunistic method significantly outperformed both standard and refined pessimistic methods. While refined pessimism is an improvement over vanilla versions, it often remains too conservative. Opportunism, by contrast, recognizes when an unobserved price with a high upper bound represents a valuable opportunity.
Conclusion
By injecting economic structure—specifically monotonicity—into the RL framework, this paper transforms a missing data problem into a geometric bounding problem. While the current model focuses on discrete price sets, it provides a vital foundation for future work involving continuous price spaces and feature-based demand. For practitioners, the "Opportunistic" approach offers a powerful tool for breaking out of historical pricing ruts and discovering more profitable strategies in data-constrained environments.
Summary updated