Tracking Consumers: The Trade-off between the Value of Granular Data and Consumers' Privacy

Abstract

In recent years, most mobile apps have started tracking consumers' location and movement patterns. This type of tracking can allow firms that access these data to better predict consumers' future behaviors and to send targeted communications. However, such tracking also raises privacy concerns among app users and regulators. This results in potential trade-offs between the value of granular tracking and privacy concerns. This paper examines three related questions. First, we examine whether granular tracking data add value when predicting consumers' retail visits relative to traditional metrics, such as consumers' demographics and past behaviors. Second, we examine whether the granularity (i.e., frequency) with which these data are tracked impacts the accuracy with which we can predict future retail visits. Finally, we examine if there is heterogeneity in the value of granularity by firm type. We address our research questions by leveraging individual-level driving data tracked via a mobile app for 31,530 consumers in Texas and their restaurant visits over 14 months between September 2018 and October 2019. We propose a machine learning (ML) framework to extract informative features from granular tracking data on consumer mobility, quantify the value of these data for predicting visits, and evaluate our model's performance under various counterfactual policies that vary the frequency with which apps can track their users. Our results show that the accuracy of prediction algorithms improves by 21% with granular tracking data relative to models that use only demographic and behavioral information on past visits. However, when tracking data are collected at longer intervals, the performance of ML algorithms decreases, but these algorithms still outperform models that use only information on demographics and past behavior. We also find that a deep learning transformer model that uses the entire sequence of latitude-longitude coordinate pairs as input outperforms the ML models by 19% in accuracy but is more computationally expensive. Our models perform significantly better with more (vs. less) granular data for non-chain rather than chain restaurants. Finally, we show an extension of our model to evaluate the app's targeting policies.

Keywords: Mobile app, location tracking, privacy, prediction, machine learning, transformers

JEL Classification: C23, C45, M1, M3, Z23, Z28