Understanding Insured Behavior through Causal Impact Analysis of IoT Streams

Advancements in Internet of Things (IoT) technologies are increasingly being leveraged by firms which incorporate Wireless Sensor Network (WSN) technology as essential components of their monitoring systems. In this paper, we focus on IoT temperature sensors deployed by insurance firms, with the expressed goal of preventing water-pipe burst (freeze loss) via real time customer alerts. In these settings, insurance firms are primarily interested in two questions: (a) Are customers responsive to alerts sent by the system? (b) Given that an alert has been sent, how long does it take for the sensor to return to “normal” behavior? To answer these questions, we propose two distinct methods. First, we develop a Gaussian process time series model to assess whether customers appreciably respond to an alert within a given time window. By comparing the model’s post-alert forecasts with observed post-alert sensor streams, we can assess customer intervention. Next, we propose a time series motif mining procedure which simultaneously (a) identifies a sensor’s typical state of behavior and (b) estimates the duration of time post-alert for the sensor to return its typical state. Ultimately, by combining the information provided by each of these methods, we can develop a more refined taxonomy of customer riskiness.