When do the stock market returns to new product preannouncements predict product performance? Empirical evidence from the U.S. automotive industry

Abstract

Substantial research has examined how stock market reactions to marketing actions affect subsequent marketing decisions. However, prior research provides limited insights into whether abnormal stock returns to a marketing action actually predict the future performance resulting from that action. This study focuses on new product preannouncements (NPPAs) and investigates the relationship between short-term stock market returns to an NPPA and the post-launch new product performance under various industry and firm conditions. Findings based on a dynamic panel data analysis of 208 NPPAs in the U.S. automotive industry between 2001 and 2014 reveal that stock returns associated with an NPPA are not an appropriate forward-looking measure of future product performance. However, under specific conditions (i.e., when the preannouncement is specific, the preannounced new product has low innovativeness, the preannouncing firm has a high reputation and invests heavily in advertising, and the preannouncement environment is less competitive), abnormal stock returns to NPPAs actually predict the future performance of new products. Thus, this study extends the marketing–finance and innovation literature with its focus on the conditions that affect the predictive power of immediate stock returns for the future performance of new products.

Appendices

Appendix 1

Appendix 2

Appendix 3

Ordinal 5-point scale for measuring innovativeness

In this study, we measure the “Innovativeness” of a model based on the amount of uncertainty and information regarding its features and future performance. For instance, if the preannounced model has already been on the market in other countries, markets will have more information about that model (e.g., its performance, reliability, features etc), compared to the available information about a “new to the world” model. That said, the “Innovativeness” of each model has been measured and coded using the scale below:

Table 12 Innovativeness scale

Appendix 4

Survival analyses using model duration in the market

We tested the robustness of our findings using another dependent variable, the entire duration of a car model in the market. Standard regression approaches are not suitable for the analysis of survival times due to right-censoring. Therefore, we used a parametric hazard model, which allows to analyze the effects of time-varying and time-constant covariates on a car model’s probability of failure. The parametric duration models assume a particular shape for the hazard rate and use a distribution (i.e., Exponential, Weibull, Lognormal, Log-Logistic, Gompertz, and Generalized Gamma) to approximate that shape. Each of these different distributions enables the estimation of a particular shape for the hazard rate (i.e. the time dependency). The precision and accuracy of the parameter estimates depend on the correct characterization of the underlying time-dependency. Therefore, it is important to determine the base hazard rate (i.e., the instantaneous probability that a model will fail at time t is constant, increasing, or decreasing with time) to examine a model’s risk of failure over time. Following Srinivasan et al. (2004), we used a multistep approach to determine the distribution that best represents the survival times of new models in the market.

First, we fitted a generalized gamma model, which has the density function below and allows for rather flexible hazard rates since it involves two shape parameters (i.e., κ and p).

$$ f(t)=\frac{\lambda p{\left(\lambda t\right)}^{p\kappa -1}{e}^{-{\left(\lambda t\right)}^p}}{\Gamma \left(\kappa \right)} $$

where

$$ {\lambda}_i={e}^{-\left({X}_i\beta \right)} $$

The generalized gamma model nests several of the other parametric models as special cases: Weibull, exponential, log-normal, and the standard gamma. Thus, it is appropriate for adjudicating between competing parametric models. The results of our estimations indicated that log-normal distribution is appropriate for our models (i.e., κ = 0) suggesting that individual hazard first rises and then declines. Second, we fit exponential, Gompertz, log-normal, log-logistic, and Weibull models separately. We estimated the exponential model, which assumes a constant hazard rate (a special case of the Weibull model, with scale parameter set to “1” and we found that this model was rejected (p < .001)). Therefore, we estimated our model using four distribution functions (Gompertz, log-normal, log-logistic, and Weibull) that accommodate monotonically and non-monotonically changing hazard rates. Although the general pattern of results was similar across models, we find that the model estimated with the log-normal hazard function fits the data better than others based on the Akaike information criterion (AIC). Therefore, we report the results of the hazard models with log-normal distribution estimated in accelerated failure time (AFT) metric with time-varying and -constant covariates and inverse Gaussian shared frailty. The survivor and density functions of the log-normal distribution in AFT metric are:

$$ S(t)=1-\Phi \left\{\frac{ \ln (t)-\upmu}{\sigma}\right\} $$

where Φ is the standard normal cumulative distribution function and. μ = Xβ .

$$ f(t)=\frac{1}{t\sigma \sqrt{2\pi }}\mathit{\exp}\left[\frac{-1}{2{\sigma}^2}{\left\{ \ln (t)-\mu \right\}}^2\right] $$

In addition to the error term that accounts for potentially observable yet omitted variables, we included a random intercept, the frailty term to control for unobserved heterogeneity, to account for factors (e.g., idiosyncratic brand characteristics) that were not included in our estimations. We used two commonly applied heterogeneity distributions, namely, Inverse Gaussian and Gamma. Based on the AIC, the Inverse Gaussian distribution outperformed the Gamma distribution, and therefore, we report the results for the Inverse Gaussian model.