PRIMER – A regression-rule learning system for intervention optimization

Abstract

We introduce intervention optimization as a new area of exploration for data mining research. Interventions are events designed to impact a corresponding time series. The task is to maximize the impact of such events by training a model on historical data. We propose PRIMER as a new regression-rule learning system for identifying sets of event features that maximize impact. PRIMER is for use when domain experts with knowledge of the intervention can specify a transfer function, or the form of the expected response in the time series. PRIMER’s objective function includes the goodness-of-fit of the average response of covered events to the transfer function. Incorporating domain knowledge in this way makes PRIMER robust to over-fitting on noise or spurious responses. PRIMER is designed to produce interpretable results, improving on the interpretability of even competing regression-rule systems for this task. It also has fewer and more intuitive parameters than competing rule-based systems. Empirically, we show that PRIMER is competitive with stateof-the-art regression techniques in a large-scale event study modeling the impact of insider trading on intra-day stock returns.