Dev log

Why that "steal of the year" probably isn't one

A bargain on a rare used car in Cyprus is exciting. It's also, statistically, usually wrong. Here's how AllCars's deal score keeps the excitement honest.

Imagine two listings. A 2018 Golf for €11,500. A rare 2012 spec sports saloon for €7,500. Both look like deals on paper. The Golf has hundreds of similar cars on the market right now. The sports saloon has maybe five — and they're all over the place.

If you blindly compare prices, both look like bargains. But there's a real difference: the Golf's "below average" actually means something. The sports saloon's "below average" might just mean there's not enough data to know what average looks like.

Pulling unreliable signals back to the middle

The fancy name for the fix is adaptive Bayesian shrinkage. The plain-English version: if a segment has hundreds of comparable cars, trust the segment. If it has a handful, don't. Pull the score back toward the wider market average — not all the way, just enough to stop the engine from screaming "Steal!" on every weird listing.

Each segment computes how noisy it is on its own (variance within) versus how different it is from neighbouring segments (variance between). The ratio drives how much it shrinks. Small, well-behaved segments barely move. Tiny, all-over-the-place segments get pulled hard.

What you see on the listing card

The deal score on a common 2018 hatchback says exactly what you expect. The deal score on a low-volume, high-variance import is more cautious — it'll still flag a real bargain when one shows up, but it stops handing out "Steal" badges to cars that just happen to be cheap because there's nothing to compare them to.

Net effect, since this shipped: dramatically fewer false-positive "Steal" badges in thin segments. Real deals in normal segments still surface as deals. You spend less time driving across the island for a "bargain" that turns out to be a rusty surprise.