Online Decisions with (Biased) Offline Data

With the growing availability of offline data, using such information to facilitate online decision making has become relevant to practice. A key challenge of using offline data, however, is that the true online data‐generating process may differ unpredictably from the offline distribution. In this paper, we study the problem of how to use available and potentially biased offline data to support online decision making for two canonical decision problems in operations management: the data-driven newsvendor, where the demand tail distribution plays a key role, and the data-driven pricing, where point estimates are critical and the decisions themselves interact with the demand distribution. We develop two online learning frameworks that incorporate offline data. The soft‐switch policy automatically balances offline information and online feedback without requiring knowledge of the bias, whereas the hard‐switch policy leverages the known extent of distributional shifts. Both learning policies achieve the optimal regret up to a logarithmic factor. We further derive explicit bounds showing how regret improves with the size and bias of the offline data.

Supplemental Material: All supplemental materials, including the code, data, and files required to reproduce the results, are available at https://doi.org/10.1287/opre.2025.2135.