We study a firm’s optimal strategy to adjust its capacity using demand information. The capacity adjustment is costly and often subject to managerial hurdles, which sometimes make it difficult to adjust capacity multiple times. To clearly analyze the impact of demand learning on the firm’s decision, we study two scenarios. In the first scenario, the firm’s capacity adjustment cost increases significantly with respect to the number of adjustments because of significant managerial hurdles, and resultantly the firm has a single opportunity to adjust capacity (single adjustment scenario). In the second scenario, the capacity adjustment cost does not change with respect to the number of adjustments because of little managerial hurdles, and therefore the firm has multiple opportunities to adjust capacity (multiple adjustment scenario). For both scenarios, we first formulate the problem as a stochastic dynamic program, and then characterize the firm’s optimal policy: when to adjust and by how much. We show that the optimal decision on when and by how much to change the capacity is not monotone in the likelihood of high demand in the single adjustment scenario, while the optimal decision is monotone under mild conditions, and the optimal policy is a control band policy in the multiple adjustment scenario. The sharp contrast reflects the impact of demand learning on the firm’s optimal capacity decision. Since computing and implementing the optimal policy is not tractable for general problems, we develop a data‐driven heuristic for each scenario. In the single adjustment scenario, we show that a two‐step heuristic, which explores demand for an appropriately chosen length of time and adjusts the capacity based on the observed demand is asymptotically optimal, and show the convergence rate. In the multiple adjustment scenario, we also show that a multistep heuristic under which the firm adjusts its capacity at a predetermined set of periods with an exponentially increasing gap between two consecutive decisions is asymptotically optimal and shows its convergence rate. We finally apply our heuristics to a numerical study and demonstrate the performance and robustness of the heuristics. The online appendix is available at https://doi.org/10.1287/opre.2016.1561.
