Research preview / source notes

What the pressure gauges mean

Raaz estimates annual country-level risk from historical observations. It does not forecast a specific location, date, or incident. Data availability and reporting vary across countries and years.

Outcomes

Model and evaluation

A regularized logistic regression learns from prior country-year observations. Inputs known at the end of each year include current and previous war status, five-year war history, real GDP per person, recent real growth, inflation, and missing-value flags. No 2025 revolution label is treated as a confirmed negative. The evaluation uses features from two years before each outcome: it trains on outcomes through 2017 and tests on later outcomes through 2025 (2024 for revolutions), then refits on all labeled years for the displayed 2027 calendar-year baseline. It uses no Raaz story text as a supervised outcome.

The benchmark always predicts the training period event frequency. A category is shown only if its later-period Brier score is at least 1% lower and the test set contains at least 25 events. The backtest is one chronological holdout, not evidence of live forecasting accuracy. Retraining and independent external evaluation are still needed.

Displayed country probabilities are raw model outputs. The band compares each probability with the historical rate among covered country-years: Calm below 0.65 times that rate; Watch from 0.65 to under 1.35 times; Elevated from 1.35 to under 2.5 times; Severe at 2.5 times and above. The global gauge is the mean of covered country estimates. It is not the probability that at least one event will happen anywhere.

Time and coverage limits

The current input snapshot ends in 2025, but its official release became available in September 2026. The 2027 number is a next-calendar-year baseline using information available by September 2026; it is not a live or rolling twelve-month forecast. One-, three-, and six-month scores remain unavailable because the annual labels cannot be divided into monthly probabilities. This page must be refreshed as source data and outcomes become available.

Raaz holds a large geolocated historical site corpus. A location and an article about an event do not by themselves create reliable country-month event labels. The corpus is useful for historical context and explanation, but the model does not currently use it as a predictive feature. Options Oracle news feeds are current snapshots rather than an archived, consistently geocoded time series; they are not yet used for training or scoring. Adding them requires a persistent archive, rights review, leakage checks, and a new chronological backtest.

Sources and reuse

• Uppsala Conflict Data Program annual armed conflict and georeferenced event datasets, via the TOCSIN CC BY 4.0 derived tables. Cite Davies et al., UCDP annual update, and Gleditsch et al. (2002), UCDP/PRIO Armed Conflict Dataset. The current source release is 26.1.
• HSE Revolution Dataset v1.1, episode start years 2000–2024. Cite the dataset and its codebook. The source workbook is not redistributed here.
• World Bank World Development Indicators, GDP per capita (constant US dollars), via TOCSIN's sourced covariates. World Bank data are CC BY 4.0.
• VIEWS pipeline core informed the separation of historical data, forecasting and evaluation. No VIEWS forecast is presented as a Raaz output.
• The globe uses Natural Earth country boundaries supplied through world-atlas. Borders are visual approximations and do not express a political position.

Code and reproduction notes are in the Raaz repository. This is a research preview and not safety, travel, financial or policy advice.