COE is a pay-as-clear auction: everyone who wins pays the same clearing price, whatever they individually bid. A bid above the clearing price therefore buys nothing. So rather than predict the price, we publish the level above which bidding has rarely been necessary — set at the 95th percentile of exercise-to-exercise moves since 2020, then tested against every following exercise.

Bidding normally runs fortnightly. When a month begins late, its first exercise follows a three-week gap — 50% more time for dealers to collect orders before bidding on their customers' behalf. Dealers have described this for years; we tested it. Since 2020, premiums rose after a three-week gap in 70–88% of exercises depending on category, against a base rate near 50%. Knowable from the calendar months ahead.

LTA announces the quota a quarter in advance, so supply is known before bidding opens. This is the one part of COE that can be stated rather than estimated. Quota per exercise is compared against the previous quarter to show whether supply is tightening or easing.

The forward figures on the Outlook tab carry today's level forward with a small drift correction. Accuracy has been about 96.6% one exercise ahead, falling to roughly 92.5% six ahead — which still means a typical miss near $8,000 on a Cat A or Cat B COE. Treat the near months as guidance and the later ones as direction only.

Trend extrapolation scored worse than assuming no change — COE mean-reverts, so momentum points the wrong way. Ten machine-learning algorithms were tested walk-forward, including gradient boosting, random forests, XGBoost and LightGBM; none beat simple carry-forward, and the best blend put 70% weight on assuming no change at all. A "wait when demand is weak" rule looked strong on past data then collapsed on data it had not seen.

GDP, CPI, unemployment, interest rates, exchange rates and property prices were each tested against next-exercise COE moves across 1,965 exercises. All correlate at less than 0.2, and GDP at almost exactly zero. Macro indicators move quarterly; COE moves fortnightly on dealer order books. We show the headline figures as background and do not pretend they predict premiums.

Every calendar month was tested against the full bidding history for each category separately, after removing each category's own underlying trend. A month is only adjusted where the direction is reliable — where the middle 50% of historical outcomes all point the same way. Only two of sixty category-month combinations passed: Cat A in March (+3.2%) and Cat D in March (+4.2%). Everything else, including Chinese New Year and year-end, showed a spread too wide to act on.

Confirmed policy quantities are used — quota changes, ARF and PARF rates. Guesses about how buyers will react to policy are not. We once modelled a demand pull-forward ahead of the EV incentive ending; tested against its only comparable episode, using commercial vehicles as a control, no pull-forward was detectable. The adjustment was removed rather than kept because it sounded plausible.