📌 What Was Introduced
A new 'variance-in-time' method uses Google Trends (GT) to detect and forecast events by collecting multiple, overlapping GT samples over time. The algorithm leverages variation in both the mean and the variance of a search term to accommodate idiosyncrasies in the GT platform.
📊 How the Method Works
- Collects multiple overlapping GT samples across time to build a richer signal than any single GT extraction.
- Uses both changes in average search volume and changes in variance to identify spikes associated with events.
- Designed to accommodate peculiarities of the GT platform that can distort single-sample estimates.
📥 Data Used to Test Forecasts
- Ground truth: protest events in the United States from the Crowd Counting Consortium, covering 2017–2019.
- A synthetic control group was created consisting of times and places where no protests occurred.
- Out-of-sample forecasts were evaluated against these true-event and control samples.
📈 Key Findings
- Out-of-sample forecasts using the variance-in-time GT method predict protests with higher accuracy than existing approaches that rely on:
- Structural predictors,
- High-frequency event data,
- Other big-data sources such as Twitter.
- The approach offers improved detection for rare but important phenomena by exploiting both mean and variance information in GT.
🔍 Why It Matters
- Provides a practical, generalizable toolkit for researchers who want to use Google Trends reliably for event forecasting.
- Advances understanding of political protest dynamics and offers a replicable alternative or complement to Twitter and structural models for short-term forecasting of rare events.
