Skip to main content
Planned. Forecast is the last layer of the Atlas family and the most demanding. It is designed but not built, and it is explicitly gated on having enough history to evaluate honestly.

​
What it is

Atlas Forecast is intended to answer what happens next between specific actors — a dyadic forecast of escalation. Not “risk in the Middle East is elevated,” but a probability attached to a named pair of actors and a defined escalation event over a defined horizon.

​
Why you should care

Almost every geopolitical forecast is unfalsifiable, and that is a choice. Vendors describe risk as “elevated” or “heightened” precisely because those words cannot be scored. A probability attached to a defined event over a defined horizon can be scored, and a vendor that publishes one is making a claim they can be held to. The commitment we are making is the scorecard, not the accuracy. Atlas Forecast will ship with a public accuracy scorecard from day one — every forecast recorded at issue time, scored with a proper scoring rule when it resolves, with the running record public whether or not it flatters us. A forecasting product without a published track record is a marketing claim. The bitemporal foundation is already in place. Honest forecast evaluation requires knowing exactly what was knowable at forecast time. Our observations are vintaged point-in-time, so an as-of query cannot see the future. That property is what makes a credible backtest possible at all, and it is why this layer is designed on top of the others rather than beside them.

​
What it is useful for

Once it exists and has a track record:
  • Early warning with a stated horizon and a stated confidence.
  • Scenario weighting — probabilities to attach to branches in a planning exercise.
  • Resource allocation across a watch list.
  • Calibration against prediction markets and internal analyst judgment.

​
Design intent

  • Dyadic and specific. Actor pair, event class, horizon, probability. No vague regional gauges.
  • Proper scoring. Brier or log score, published, with a naive base-rate benchmark alongside — a forecast that cannot beat the base rate is worth knowing about.
  • No look-ahead, structurally. Forecasts are generated from vintaged as-of data, so an evaluation cannot leak future information.
  • Grounded in the layers below. GPR supplies deviation, Posture supplies condition, CNP supplies capability. Forecast is a function of those, not a separate black box.

​
The honest blocker

We do not have enough history to build this yet, and we will not pretend otherwise.Consistently coded events begin in March 2026. Escalation forecasting requires many observed escalation episodes across many dyads to estimate and validate anything, and a few months of history contains too few. Building it now would produce a model we could not honestly evaluate — which is exactly the failure mode the scorecard exists to prevent.Forecast is therefore gated on accrued history rather than on engineering effort. It ships when it can be scored, not before.

​
Influences and references

Political Instability Task Force — Goldstone et al. (2010)

The benchmark demonstration that political instability is forecastable at useful accuracy from a small, interpretable variable set — and the standard against which any new instability forecast should be measured.

ICEWS — Integrated Crisis Early Warning System

The DARPA-originated event-data early-warning program: coded political events used for forecasting. The clearest precedent for forecasting from an automated event stream, and a cautionary tale about coding quality propagating into forecasts.

The Good Judgment Project — Tetlock et al.

Established that forecasting skill is real, measurable with proper scoring rules, and that calibration must be demonstrated rather than assumed.

Uppsala Conflict Data Program / ViEWS

The ViEWS political violence early-warning system — a serious, publicly evaluated conflict forecasting effort, and the most direct methodological comparator for this layer.

​
Status

Not started, and deliberately last. Sequenced after GPR reaches general availability, Posture ships, and enough consistently coded history has accrued to support an honest evaluation.