Model one latent path
Represent the evolving fire regime with one persistent state process rather than unrelated horizon-specific predictors.
← All researchMF-PRISM-FIRE-2026-02 / interactive guide
THE IDEA, MADE VISIBLE
One future. Several horizons.
A coherent first-event forecast grows along one path. Missing outcomes must remain visible in evaluation.
Research reportObservation-Aware Wildfire Regime Forecasting
Use one persistent path model
This toy process has a constant probability h of a first event during each hour, conditional on no earlier event. The same process determines every horizon.
- 3-hour model risk
- 31.9%
- 12-hour model risk
- 78.4%
- Batch event fraction
- 25–58%
One synthetic path gives 31.9% by 3 hours and 78.4% by 12. Separately, 4 unresolved labels allow a batch event fraction of 25–58%.
P(first event by t) = 1 − (1−h)ᵗ
TRY THIS
Increase unresolved outcomes. The admissible event fraction widens, while the coherent model curve remains unchanged.
What this experiment represents. Constant-hazard synthetic first-event process and a separate constructed label-completeness example. The label fraction is not a calibration test of this hazard curve. No operational probabilities are estimated.
Wildfire Intelligence · Research Paper
Observation-Aware Wildfire Regime Forecasting
Coherent multi-horizon risk, unresolved outcomes, and incident-balanced selective validation
Research reportMF-PRISM-FIRE-2026-02 / v0.2
Develop coherent risk statements across several forecast horizons while retaining uncertainty about the latent state and unresolved outcomes. The evidence remains conditional and synthetic, not a calibrated operational wildfire forecast.
- Identifier
- MF-PRISM-FIRE-2026-02
- Series
- Wildfire Intelligence
- Edition
- Version 0.2
- Length
- 25 pages
- Reserved DOI
- 10.5281/zenodo.22726324 (record reserved)
01
Plain-language guide
3 minute explanation
What this paper is really saying.
The paper makes short-, medium-, and longer-horizon wildfire risk statements come from one persistent model while retaining unresolved outcomes instead of quietly deleting them.
wildfireforecastingselective evaluation
A forecasting system can look precise while contradicting itself: the 6-hour risk may be high, the 12-hour risk low, and the 24-hour risk high again because separate models were fit independently.
This paper proposes one latent regime path that generates all horizons coherently. It also treats unresolved incidents as unresolved, supports selective abstention, and balances evaluation by incident so large or well-observed fires do not dominate the score.
02
How it works
The argument, without the notation.
The paper’s technical details matter, but the basic route can be understood in three moves.
Derive every horizon coherently
Compute multi-horizon risk from the same path distribution so probabilities obey a common internal model.
Evaluate without hiding uncertainty
Retain unresolved labels, report selective coverage, and aggregate results in an incident-balanced way.
03
What it adds
The useful takeaways.
Multi-horizon probabilities are constrained to be mutually coherent under one model.
Unresolved outcomes are retained rather than silently recoded or discarded.
Selective forecasting is evaluated together with coverage, so abstention cannot masquerade as accuracy.
04
Scope & limits
What this does—and does not—establish.
Current status
Conditional forecasting and selective-evaluation methods with synthetic evidence. No operational forecast or calibrated real-fire probabilities are established. Research only; not an official emergency warning.
This report develops a conditional method and evaluates it within the stated evidence. It does not establish production performance, operational safety, or calibrated real-world predictive skill beyond that evidence.
The probabilities are not calibrated for operational real-fire use.
A coherent model can still be wrong if its regime assumptions are misspecified.
This research is not an emergency warning system.
05
Why it matters
Why anyone should care.
Operational-looking probabilities are dangerous when horizons conflict or labels are incomplete. The framework forces one persistent path model and evaluation that does not silently discard unresolved incidents.
06
Key terms
The vocabulary, decoded.
These are the terms needed to understand the claim. The full paper uses them more precisely.
Forecast horizon
The future time interval for which a prediction is issued.
Latent state
An unobserved condition inferred indirectly from available measurements.
Selective prediction
A system that may abstain when its evidence is insufficient.
Calibration
Agreement between stated probabilities and observed frequencies over comparable cases.
07
Independent review
Where scrutiny should concentrate.
The strongest review is not a general reaction. It tests the steps most capable of changing the conclusion.
- 01One persistent model over the full forecast path.
- 02Unresolved outcome labels and incident-balanced evaluation.
- 03Selective evaluation and model misspecification boundaries.
Publication record
Go from explanation to evidence.
This page is a reading guide, not a substitute for the manuscript. The public record links the explanation to the paper, source package, review materials, and persistent identifier.
- Document
- Research Paper
- Review status
- Open for independent review
- Published
- 2026-09-12
- DOI status
- Reserved · 10.5281/zenodo.22726324
- Canonical record
- MF-PRISM-FIRE-2026-02
Metriq PRISM Laboratory, Observation-Aware Wildfire Regime Forecasting, Metriq PRISM Laboratory Research Paper MF-PRISM-FIRE-2026-02, Version 0.2, 2026. Corresponding contributor: Daniel H. Jeffery, ORCID 0009-0001-1200-6042. DOI: 10.5281/zenodo.22726324.