Methodology

There is no single “most accurate” forecasting model for e-commerce demand. Accuracy is a property of a model and the data it is given, and the data here changes shape as sources come on. So the method is chosen by what the evidence can support, and every method publishes the same thing: a distribution over units, as a quantile grid, never a point estimate.

What runs today: pooled comparables

The corpus is cross-sectional snapshots — a listing’s age, price, licence and verified-buyer review count — with no receipts and few long series. That is a cold-start problem, and the method that works for cold starts is to borrow strength from comparables. Each niche fits a Gamma–Poisson model to the review rate of physical listings in its cohort, with an empirical-Bayes prior from the category, and projects a new entrant’s units over the horizon. Review counts are used because a review requires a purchase; the vendor’s own monthly-sales estimate is shown beside the result for comparison and never fed into it.

What turns on with sold-history: quantile gradient boosting

Once eBay sold listings are ingested there are real transactions with dates and prices. At that point the strongest documented approach for retail demand is a single global model trained across every product at once — gradient-boosted trees fitted with pinball loss at each published quantile — over features the corpus already carries: price band, category, listing age, review velocity, physical-versus-digital, seasonality, and relative search interest when Trends is on. This is the family that dominated the M5 retail forecasting competition, and it wins precisely because it learns across thousands of thin series rather than fitting each alone. Conformal calibration on held-out outcomes then adjusts the intervals so the stated coverage is the real one.

Deep sequence models are not on this list yet. They need thousands of long series to beat boosted trees, and until the index has them, they would be a more expensive way to be less well calibrated.

How it improves: the forecast ledger

Every published forecast is written down as it was published — the full quantile grid, the interval, the model version — and when its horizon closes the observed outcome is written beside it, tagged with how it was established: seller receipts, a marketplace report, or a review delta. The pairs are scored on coverage, the probability integral transform and CRPS. If an 80% interval is catching 60% of outcomes the report says so and by how much to widen; if it is catching 95% it says the model is being timid. Model versions are compared on the same ledger, so “more accurate” is a back-tested number rather than a claim.

The ledger is append-only. A forecast cannot be edited after its outcome is known; a correction is a new row. That is the one property the whole loop depends on.