v1.1 · framework
A forecast is only as honest as its distance from the data.
Most "accountability" frameworks for predictions work backward: a surprising result happens, a new rule gets bolted on, and after a few rounds the system looks rigorous but is really just overfitting to a handful of matches.
FORECAST/ORIGIN works the other direction. Build the model from real distributions first. Let evidence move it by an amount proportional to how much evidence actually exists. No rule exists because of one match. No floor exists because something "deserves" it.
Stage 1
Data intake (no judgment yet)
Goals for/against per 90, season + last-5 separately. Table position. Home/away split. H2H, capped low unless n > 5. If a stat isn't available, the report says "unavailable" — it is never quietly filled with a league-average placeholder.
Stage 2
Model, not narrative
A Poisson / Dixon-Coles-style expected-goals model. The full scoreline matrix is generated mathematically — every cell is Poisson-derived, not a manually set floor. Recent form is blended at a fixed, stated weight (70% season / 30% last-5), chosen in advance rather than per-match to fit a preferred conclusion.
Stage 3
Evidence-proportional updates
Bayesian-style nudges sized to actual sample size. n=1 moves the model a little; n=5 moves it a lot. The report says what n is. "One data point" and "consistent five-match pattern" never get the same-sized bump.
Stage 3.5
AI-intuition layer (new in v1.1)
The model output is locked first. Only then does an LLM propose qualitative reads and numeric nudges. Hard rules:
- Every nudge is capped at ±5 percentage points. Absolute.
- Every nudge is a sentence a reader could disagree with.
- The qualitative read and the numeric nudge are reported separately.
- Delete the whole intuition section and a complete forecast still stands underneath.
- If there is no genuine read, the section says "no material adjustment" rather than inventing one.
Stage 4
Uncertainty is reported, not decorated
Intervals come from 10,000-iteration Poisson simulation, not from a stylistic "give it a 5-point spread" habit. If two outcomes are genuinely close, the report says so.
Stage 5
Calibration (earned, not claimed)
Every published probability is tracked against the actual outcome. After 30+ forecasts, the Brier score is computed. Until then, every forecast carries the line: "This model has not yet been back-tested against a real sample. Treat confidence levels as provisional."