Counter-Factual Hidden Event Detection: A Proof of Concept Using Machine Learning in Time Series

We propose a new framework for detecting hidden events based on historical data from previous events and related time series data that are known (or presumed) to be affected by the event. In particular, a forecasting model of the time-series data is trained using known occurrences of the hidden events. Then, for a given time interval where a hidden event may have occurred, the time series is forecasted under various hypotheses, including the hidden event happening (possibly at various time intervals) and not happening. The resulting forecasts are compared with the actual time series for the interval where the event would have had its maximum effect, and the hypothesis that leads to a forecast closest to the actual data is selected. The methodology is evaluated under two scenarios. The performance of the methodology is satisfactory in both cases, showing its potential for hidden event detection.