Every prediction we publish is stored before the exercise opens and scored against the result once LTA publishes it. Nothing is removed after the fact. Where the model is weak, it says so.
Accuracy by distance
Measured against LTA bidding results across 2025–2026. Accuracy falls the further ahead you look, which is why the outlook on the main page is labelled indicative rather than a forecast.
Scored predictions
Exercise
Category
Predicted
Actual
Off by
Accuracy
loading…
Six-month outlook, scored by month
The table above scores each bidding exercise. This one scores the six-month outlook against the monthly average, which is how it was originally tracked. Both are shown; neither is adjusted after the fact.
Loading accuracy data...
What we tested and rejected
Building this tool meant discarding more than we kept. Recorded here so the same ground is not covered twice:
Trend extrapolation — scored worse than assuming no change. COE mean-reverts; momentum correlates negatively with the next move.
Machine learning — ten algorithms tested walk-forward (ridge, gradient boosting, random forest, XGBoost, LightGBM and others). None beat carry-forward. The best blend put 70% weight on simply assuming no change.
Macro indicators — GDP, CPI, unemployment, interest rates, exchange rates and property prices all correlate with next-exercise moves at less than 0.2. Macro moves quarterly; COE moves fortnightly on dealer order books.
“Wait when demand is weak” — looked strong on past data, then collapsed on data it had not seen. Discarded.
What survived
The bid ceiling — covered 96–100% of exercises since 2020, tested walk-forward. It works because COE is a pay-as-clear auction: bidding above the clearing price buys nothing.
The three-week gap — when three weeks pass between exercises instead of two, dealers collect more orders and premiums have risen 70–88% of the time depending on category. Knowable from the calendar months ahead.
Published quota — LTA announces supply a quarter in advance, so it is certainty rather than estimate.