Abstract

Product-harm events are always a nightmare for all stakeholders. Analysts believe that defective items may not only provide risks to the general population, but can likewise cause critical monetary and reputational harm to the firms. Since ignoring a problem does not lead to having it go away, more research is needed to shed new light on the way crisis and risk communication should take place once necessary. Prior study has suggested the complexities of consumer word of mouth effects and how to accurately forecast the impacts of product-harm events on firm value as important subjects. This study extracts the sentiments of consumer complaints in the context of product defects and examines if including consumer sentiment in time series models can improve forecasting performance. Authors make an empirical comparison between two multivariate time series forecasting methods: VAR (vector autoregressive model), and deep learning LSTM (long short-term memory model). Unique datasets, containing five-year data of all automobile nameplates for three major manufacturers in the U.S. are analyzed. The one-step rolling forecast approach is applied to validate time series forecasting values. The results of mean RMSE suggest that LSTM outperforms VAR predictive ability of firm value, and on average obtains 59.02% reduction in error rates when compared with error rates of VAR. It is also noticed that adding consumer sentiment in modeling can improve the predictive performance of both LSTM and VAR models; however, VAR-based models make greater progress in predictive error reduction with consumer sentiment. Implications for marketing research and managerial contributions are discussed.