Most weather products tell you the forecast. Forecast Fork tells you whether to trust it — which competing scenarios are still on the table, whether each one is gaining or losing probability run over run, and what it would be worth if it won. It is the open, continuously-verified instability layer of Weather Trader — and every number below is pulled from it live, as you load this page.
See the scenarios live → How it's verifiedloading live from weathertrader.com…
This is not a mock-up. It is fetched, as you load the page, from Weather Trader's verification engine: bin every run by the ensemble spread it showed at issue time, plot the skill it actually verified. The down-slope is the proof that spread is a bust signal — known before the outcome. That is the Forecast Fork.
A fork is not a metaphor here. Every member of every ensemble — ECMWF EPS, GEFS, AIFS-ENS and all 64 of Google's WeatherNext 2 — is classified into the same named regimes at every lead, and the shares below are those counts, live from the newest cycle. Where the lines split, the atmosphere's futures split.
loading the member distribution…
Read it left to right: near the start the members agree and one regime holds all the probability; downstream the shares divide. The lead where they cross is the fork — and it is known, member by member, the moment the cycle is issued.
Anyone can cluster an ensemble once. The hard part — and the whole point — is that a scenario must still be the same scenario next run. Every cycle, Weather Trader clusters the pooled members into competing scenarios, then matches this run's clusters to the last run's by area-weighted pattern correlation with an exact one-to-one assignment. Storylines therefore persist: they are born, they gain or lose probability mass, and they die.
loading the scenario tracks…
These are the engine's own composites, rendered on demand — the ensemble mean is the no-clustering baseline, and each scenario is what the atmosphere looks like if that branch wins. Pick your own pool, region, lead and number of scenarios — or build a custom pool from any set of shipped ensembles — on weathertrader.com/clusters.
A scenario is only interesting if it changes something you can act on. Each branch above is carried through to a downstream quantity: population-weighted degree days over the Lower 48, accumulated across the five-day window, every day scored against its own climatology. Same members, same window, one number per future — plus the probability-weighted expectation and the spread the branches imply.
| loading the scenario book… |
This is the Fork made spendable. A forecast that is "uncertain" is a sentence; a forecast whose branches differ by — cooling degree days over a five-day window is a position size. The full book — four pools, three windows, every k, with each scenario's biggest station movers — is on weathertrader.com/energy.
A consensus is only worth what its independence is worth. We measure it directly: correlate every model's error field against the same analysis, and count the effective number of independent forecasts. Eight models that all miss the same way are one model with error bars.
loading the diversity matrix…
This is why the scenario engine reports each branch's membership by system. A scenario carried by one model family is a different object from the same scenario carried by all of them — and an A.I. model that echoes its IFS/ERA5 training is not an independent vote. Full board: weathertrader.com/herding.
A forecast can be confident — every model and member converging on one outcome — or forked: split into distinct scenarios the atmosphere hasn't chosen between yet. The second kind busts. We keep the three ways of saying that separate, and label each with what has actually been proven about it.
Normalized ensemble disagreement per region × lead, fully known at issue time. Its association with realized skill is the chart at the top of this page — measured, not asserted.
A fitted logistic P(ACC < threshold), scored two ways: leave-init-out, and strict walk-forward — past-only, the honest deployment metric. Reliability is published with it.
The Revision Tensor: run-over-run RMS change of the 500 hPa outlook for the same valid day, by lead and region. Temporal instability — distinct from spread, and from P(bust).
We combine the three into one headline index only once each is independently, out-of-sample calibrated — not before. Method & live component status on weathertrader.com/method.
Treat the models as bettors, each staking its Day-5 field. A bookmaker with hindsight — who knows the verifying analysis — would set each run's weights to the blend that minimises the error that actually happened. That book cannot be run in real time. But it can be computed afterwards, and the distance between it and the consensus we actually publish is the information left on the table.
loading the hindsight ledger…
Forecast Fork is not ensemble spread. For a fixed valid time the ensemble is a mixture of coherent scenarios, and we track those branches across successive model runs — not re-clustering and forgetting each cycle. The object is a branching graph: how forecast alternatives emerge, split, merge, and gain or lose probability as the forecast evolves toward reality. The branching graph above is that equation, running.
The system runs in both directions. Forward Fork: from the current state, which outcomes are possible and how are their probabilities evolving? Reverse Fork: given an outcome — the analysis that verified, or a station's daily high above a threshold — which branches carry it, what must stay true, and what would kill it? The branch shares at issue are the prior, each branch's fit to the verifying analysis is the evidence, and Bayes gives the posterior: P(Bk|E) ∝ πk·qk(E) — live on weathertrader.com/reverse.
The current branches — how many, how separated, their probability, and their membership by system. Four pools (EPS · full physics · A.I. · the Super Ensemble), five regions, seven leads and windows, k = 2–4, every combination published each run. /clusters
Scenarios matched cycle to cycle by pattern correlation with exact assignment — births, deaths, and the run-over-run share trajectory of every storyline. The branching graph itself.
Which branch is gaining, losing, stabilizing or reversing — Δπ read straight off the tracks, not re-derived each cycle from scratch.
Every branch propagated to a downstream outcome: accumulated HDD/CDD on the census population weights, the stations that move most, the probability-weighted expectation and the spread the branches imply. /energy
The effective number of independent model families — from the pairwise correlation of error fields, plus each scenario's system composition. A.I. models echo their IFS/ERA5 training; this says by how much. /herding
Given an outcome, the branches that carried it: prior share at issue, the evidence, and the Bayes posterior — with the trail of how that branch's probability evolved before it won. /reverse
The Revision Tensor — run-over-run RMS change of the outlook for the same valid day, by lead and region, with the current cycle's read against the archive's normal. /verify
A fitted, out-of-sample-scored probability the forecast fails — leave-init-out and strict walk-forward, with the reliability table published. Not merely spread, and not yet promoted.
After verification: when the winning branch first appeared, how its probability evolved, and what was knowable when. Written up weekly in the Casebook from the frozen record.
Status is honest: live is running publicly and refreshed every cycle, experimental is built and being out-of-sample verified. All of it runs on our own multi-terabyte ensemble archive, computed near the data — member fields reduced on the machines that hold them, so a pooled clustering of two hundred members is a few megabytes on the wire, not a few terabytes. Full method and component status: weathertrader.com/method.
Member fields from every ensemble reduced to a common grid, so EPS, GEFS, AIFS-ENS, WeatherNext 2, HRES, AIFS and GFS can be pooled and clustered as one population.
One number per region & lead — how forked the forecast is — on the Forecast desk's agreement heat.
Competing patterns cut from the pooled members, each with its own 500 hPa composite, temperature anomaly and degree-day consequence. Build your own on the Cluster Maker.
Spread↔skill correlation and a walk-forward-scored bust model, published and continuously updated — the charts above. See the Method.
The synoptic regime the flow is in — blocked, zonal, PNA− — and which model owns it. The named reason a forecast is unstable.
Mean anomaly correlation at Day 5/7/10 over every run that verified in the last week, month or half-year — all models against one common analysis. /verify
Tamper-evident, content-addressed record of every forecast as issued — the sha256 is the identity, so any change is detectable. Hash-chained + externally-anchored roots are on the roadmap.
Forecast Fork is the transparent, auditable layer of Weather Trader — a prediction market for the atmosphere. The maps and verification are free and public; the goal is for "the Forecast Fork is high" to become how people say a forecast is unsettled. Start on the live scenario engine or read the method.