Predicting the Unpredictable: Reproducible BiLSTM Forecasting of Incident Counts in the Global Terrorism Database (GTD)

Summary

This research focuses on predicting the number of terrorist incidents each week, using data from 1970 to 2016. The goal is to see if advanced machine learning can accurately forecast these events, and provide a reliable starting point for future research.

What's the problem?

Predicting terrorism is really hard because it's complex and influenced by many factors. Existing methods often aren't very accurate, and it's difficult to compare different approaches because research isn't always shared openly. The challenge is to build a forecasting model that's better than simpler methods and is also transparent so others can build upon it.

What's the solution?

The researchers built a forecasting system using a type of neural network called a Bidirectional LSTM, which is good at understanding patterns in sequences of data. They compared this to simpler forecasting methods like looking at past trends and using a standard statistical model called ARIMA. They carefully set up the data to ensure fair comparisons and tested their model on data it hadn't seen before. They also experimented with different ways to feed information into the model, like varying the amount of historical data used and the features included, to see what worked best.

Why it matters?

This work is important because it provides a strong, publicly available baseline for forecasting terrorism incidents. The BiLSTM model they developed outperformed other methods, meaning it's more accurate at predicting weekly incident counts. By sharing their code and data, they're helping other researchers improve these forecasts and potentially contribute to efforts to understand and prevent terrorism. The study also highlights which factors seem to be most important in predicting incidents, like past incident counts.