How prices are calculated
Every market price on Prayk, the “Good deal” chip, the Deal Score and the valuation tools run on the same open methodology. Here are the sources, the rules and the live measured accuracy.
Data sources
Prices are based on real asking-price listings collected from Azerbaijani classifieds sites and on Prayk's own catalogue. Each listing is one observation: price, parameters, location, when it appeared and when it disappeared.
- Observations, last 90 days
- 195 492
- Price baskets
- 59 083
- Displayable baskets (N ≥ 8)
- 20 176
- Last update
- 28 September 2026
How a price is computed
- 1Observations are split into tiered baskets: cars — make → model → year (±1) → body → engine/gearbox/mileage; property — deal → type → district → rooms → area band → building type.
- 2In each basket outliers are dropped (IQR × 1,5), then the median and the P25–P75 range are computed. Median, not average: a few inflated listings cannot drag the result.
- 3Reading starts from the most precise basket and stops at the first one with at least 8 listings. Baskets are never summed — a single basket is always read.
- 4For a property listing with coordinates a geo-basket is built from neighbours within 500 m — more precise than district statistics.
- 5Baskets are rebuilt every night; the market gap (%) on a listing card refreshes every 6 hours.
The role of AI
The price does not come from a chat model (ChatGPT-style) — it is computed by statistical baskets built on real listings and, in the valuation tools, by a hedonic model. AI only checks a listing's text and photos, explains risks and puts the computed result into plain language; it never changes the number.
Honesty rules
- A basket with N < 8 is never shown — a “market price” from a handful of samples is misleading.
- No score is given at the shallowest (category- or brand-only) basket: a small device is not compared with its whole category.
- Listings imported from other sites get no “new seller” penalty — recognition is when the bot first saw it, not the seller's age.
- A price 45 % or more below market is not “excellent”, it is a RISK signal — bait prices are capped by the score ceiling.
- All prices are market ASKING prices — not actual transaction prices.
- If at least 3 similar listings in the same building/complex (150 m) disagree with the neighbourhood verdict, we don't say cheap/expensive — the area median mixes premium and ordinary buildings.
Accuracy (backtest)
The valuation model is checked every week on listings it has not seen: median absolute percentage error (MdAPE) and the share of predictions within ±10 % / ±20 %.
| Segment | Confidence | N | Coverage | MdAPE | ±10 % | ±20 % |
|---|---|---|---|---|---|---|
| All segments | All | 2 014 | 96,2 % | 12,1 % | 43,8 % | 69,0 % |
| All segments | High | 1 514 | 100,0 % | 10,4 % | 48,8 % | 75,6 % |
| All segments | Low | 124 | 38,7 % | 36,0 % | 16,7 % | 31,3 % |
| All segments | Medium | 376 | 100,0 % | 22,1 % | 27,4 % | 47,1 % |
| auto · * | All | 2 062 | 94,5 % | 9,6 % | 51,8 % | 78,1 % |
| auto · * | High | 1 834 | 100,0 % | 9,4 % | 52,8 % | 79,5 % |
| auto · * | Low | 123 | 7,3 % | 15,7 % | 33,3 % | 66,7 % |
| auto · * | Medium | 105 | 100,0 % | 16,8 % | 34,3 % | 55,2 % |
Backtest date: 2026-09-28
Update cadence
Baskets are rebuilt nightly; card deltas every 6 hours; valuation coefficients nightly; the backtest weekly.
This page is updated whenever the methodology changes. Figures come from the live database and are for information only.