This paper introduces a way of testing the predictive power of structural models. Based on retail market scanner data for ground coffee in Germany during the period 2002 to 2012, I compare marginal cost estimates for coffee store brands from an out-ofsample prediction to estimates based on a structural model of retail competition. The comparison reveals substantial differences. While the marginal cost estimates using an out-of-sample prediction closely follow the world market prices for coffee beans representing the main driver for changes in the cost of ground coffee, the estimates based on a structural model are considerably higher following a cartel-induced increase in wholesale prices for national brands. Given that the cost of coffee store brands are not affected by the cartel-induced increase in wholesale prices, the increase in marginal costs, based on the structural model, reflects the model’s failure to capture retailers’ pricing choices in a multiproduct environment.