How does it work?

A short guide to the escalation probability tool and how to read the results.

What is escalation probability?

Escalation probability is an estimate of the likelihood that a conflict or dispute between two countries could escalate (e.g. toward more serious confrontation or military action). The values you see are derived from a model and are for informational purposes.

How to use the map

Hover over a country to see its name.
Click a first country to select it (it will be highlighted).
Click a second, different country to see the escalation probability between the two.
A popup will show the probability, risk level, a short interpretation, and an explanation of how the score was computed (feature contributions).
Clicking a third country clears the previous pair and starts a new selection.

Risk levels

The popup assigns a simple risk label based on the probability:

Low risk — probability below 15%
Medium risk — between 15% and 40%
High risk — above 40%

What the numbers mean

Probabilities are shown as a percentage (e.g. 23.07%). When possible, the tool also displays a “1 in X” formulation (e.g. “Roughly 1 in 4 chance”) to make the likelihood easier to grasp. For each pair, the popup includes a “Why this probability?” section with a short explanation: which features (news tone, distance, military spending, NATO, etc.) contributed and how.

How the probability is calculated (math)

For each country pair (A, B), the model outputs a probability P in [0, 1]. It works in two steps.

1. Linear combination (logit) z = b + w₁x₁ + w₂x₂ + \dots + w₉x₉ Here b is the bias (intercept) and each xₖ is a feature in [0, 1]; wₖ is its weight. So z is a real number (the “logit”).

2. Sigmoid (probability) P = σ(z) = 1 / (1 + e -z) The sigmoid σ(z) squashes z into (0, 1). Higher z \to higher P (more escalation risk); lower z \to lower P .

The contribution of feature k to the logit is cₖ = wₖ·xₖ. So z = b + c₁ + … + c₉. The popup shows each cₖ so you can see which factors push the score up or down.

The 9 features (all in [0, 1]) are: news negativity, news intensity, escalation/conflict events (GDELT); contiguity, distance closeness, common language (CEPII); both in NATO; military expenditure (mean) (World Bank); recent interstate conflict (UCDP). For scaling rules and full details, see readme.md.

About the data

Probabilities and explanations are loaded from prob_explained.csv, which lists each country pair with its probability and a text justification (logit, bias, and per-feature contributions). The map uses standard country names; some names may differ slightly between the map and the CSV, but the tool matches them where possible.

This tool is for exploration and education only. The underlying model and data have limitations; do not rely on it for policy or real-world decisions.