Sales Forecasting of New Entertainment Media Products

ABSTRACT

Managers predict the sales of new entertainment products prior to their release using comparables, such as similar books from the same author or movies with the same actors. In this study, the authors analyze whether diffusion models for media products provide helpful support in the management task of predicting prelaunch sales of the first distribution channel for three different product categories. They compare the performance of predictions based on (a) simple success factor regressions (OLS) and (b) diffusion models against real management predictions. Based on samples covering the German music, film, and the literary market, we show that model-based forecasts outperform the forecasts of management teams for the majority of the products. In contrast, management is superior in forecasting top sellers. This is due to unobserved factors arising from more management attention attached towards super stars. The authors do not find substantial prediction differences between simple success factor regressions and more complex diffusion models. Thus, managers interested in total sales estimates can easily rely on OLS based success factor predictions. Advertising and product differentiation factors with respect to quality (e.g., star power, critics, or country of origin) are across all 3 industries highly relevant for sales predictions, whereas others variables (e.g., price, distribution power, season, or competition) differ across industries.

Acknowledgments

We thank the anonymous reviewers and the editors for their valuable feedback to previous versions of this manuscript.

Notes

¹ We use the term sleeper movie (and not platform release) as the term better reflects the demand side of the diffusion, while the platform release refers more to the strategy of the distributor.

² The sales number provided by Media Control was multiplied by the factor 100/85 to represent the entire market.

³ Compared to prior research, the results of fitting the diffusion shapes are satisfying, with average R² values for movies of 98.73% (Lehmann & Weinberg, 2000 for movies: 93%). For music and books, we achieve similar results, with values of 96.38% for music and 83.81% for books. Beck (2007) reported R² values of 89.93% for only four books (median R²: 81.25%). The fit of the diffusion shapes with other models from the motion picture industry (Generalized Gamma, Exponential) and with the Bass model are inferior compared to the Weibull model, although the performance is satisfactory and the results do not substantially differ (tests of mean differences revealed no significant differences between all models). The diffusion of books is, in general, more difficult to capture due to strong weekly sales fluctuations and longer life cycles.

⁴ The answers of the 12 moviegoers and two experts significantly correlate with r = .84 (major actors) and r = .88 (director), respectively. Measuring starpower of actors for movies that are running in German theatres is a very difficult task. Because German movies include national and international stars, we cannot rely only on starpower measures that focus on the U.S. actors (e.g., IMDb’s starmeter). Nevertheless, we correlated our variable of the questionnaire with the IMDb starmeter (Hofmann et al., 2017) only for a subset of U.S. movies. The rank correlation coefficient of Spearman is r = .47 and highly significant.

⁵ Often albums are released in a variety of editions (e.g., with additional songs or extras). This has to be accounted for because the market potential of an album is split between those editions.

⁶ We estimated the amount of screens using OLS following an instrumental variable approach: In a first stage regression we used all success factors plus the distribution power as an instrumental variable to explain the number of screens. Then, the number of screens was predicted using the first stage regression results. Instead of using the predicted amount of screens (SCR_PRED) for the second stage regression, we follow the control function approach and use the residuals to minimize multi-collinearity: SCR_PRED-SCREENS = SCR_RES.

⁷ We have tested various versions of modeling the respective parameters (e.g., multiplicative models), which have not increased the overall forecasting performance. Thus, we follow the modeling approach by Lee et al. (2003).

⁸ Of course, the R² can be inflated by adding further variables to the model.

⁹ If a variable is not used for the forecast, we adjusted the constant by a correction term: constant – (coefficient mean of the variable).

¹⁰ Selling category defines the categories from Flop to Top (movie has six categories from 1 = Flop to 6 = Top, music and books have each four categories from 1 = Flop to 4 = Top).

¹¹ To control for potential effect size heterogeneity, we tested whether the forecasts can be improved by running a bivariate logistic regression that separates the lowest selling category (=0) from the other categories (=1). Referring to the logit model, we run two regression models, each per category (Tables A6 and A8 in the appendix). Therefore, the forecast of a new movie is first a forecast of the category by the logit model and then a forecast by the respective regression function within this category.