More information, better forecasting? Empirical evidence from integrating customer sentiment and Google Trends

Our study delves into a major – and yet a largely unanswered research question in the forecasting literature: does the integration of additional information into forecasting models improve forecasting performance? We experiment with incorporating customer sentiment and/or auxiliary data (Google Trends) into both statistical and advanced machine-learning forecasting models; the latter is used to examine whether this additional information is better utilised linearly or non-linearly. This is a methodological contribution that focuses on the added value of additional information; it is not a paper on the performance of specific models or methods, which are used in this research only as representatives of their class, predominantly to capture nonlinearities. The empirical setting involves forecasting sales of gaming consoles from Nintendo, Sony PlayStation, and Microsoft Xbox, using customer reviews on Amazon and Google Trends search trend data. We compared the sentiment-extraction tools VADER and TextBlob, and our results favoured the former; thus, the choice of tool matters. The most important finding was that forecasting sales using Google Trends and customer-review-related variables generally increases prediction accuracy, and there is preliminary evidence that this additional information is better utilised by non-linear models (i.e., Random Forests).