Online Decisions with (Biased) Offline Data

Making Better Decisions with Imperfect Historical Data Historical data can provide valuable information for online decision making, but its usefulness may be limited when the environment has changed. In “Online Decisions with (Biased) Offline Data,” Ahn, Zhang, and Zhang study how firms can combine potentially biased offline data with newly collected online observations in two canonical operations management problems: dynamic inventory decisions and dynamic pricing. The authors develop two online learning policies that balance the benefits of historical information against the risk of relying on data from a different distribution. The soft-switch policy automatically adjusts the reliance on offline and online data without requiring prior knowledge of the degree of bias. The hard-switch policy instead uses information about the magnitude of the distributional shift to determine when to rely primarily on online data. Both policies achieve optimal regret up to a logarithmic factor, whereas the analysis quantifies how the value of offline data depends on its size and bias.