While numerous researchers have proposed different models to forecast trial sales for new products, there is little systematic understanding about which of these models works best, and under what circumstances these findings change. In this paper, we provide a comprehensive investigation of eight leading published models and three different parameter estimation methods. Across 19 different datasets encompassing a variety of consumer packaged goods, we observe several systematic patterns that link differences in model specification and estimation to forecasting accuracy. Major findings include the following observations: (1) when dealing with consumer packaged goods, simple models that allow for relatively limited flexibility (e.g. no S‐shaped curves) in the calibration period provide significantly better forecasts than more complex specifications; (2) models that explicitly accommodate heterogeneity in purchasing rates across consumers tend to offer better forecasts than those that do not; and (3) maximum likelihood estimation appears to offer more accurate and stable forecasts than non‐linear least squares. We elaborate on these and other findings, and offer suggested directions for future research in this area. © 1998 John Wiley & Sons, Ltd.