Israel Used-Car Deal Finder

Find used cars on Yad2 that are priced below what comparable cars are listed for right now. The actor pulls a pool of live Yad2 car listings, groups them into cohorts of similar cars (same manufacturer, model, and year, adjusted for mileage), computes a robust fair market price per cohort, and flags any listing priced meaningfully below it.

The pitch is simple: priced below what similar cars are actually listed for right now. No external price guide, no guesswork, only the live Yad2 market.

What it does

1. Pulls a comparison pool of Yad2 car listings (the market baseline).
2. Groups them into cohorts by manufacturer, model, and year, with Hebrew-aware name matching.
3. Computes a robust fair market price per cohort using a trimmed median, and adjusts for mileage and model year.
4. Compares each listing's asking price to its cohort's fair value.
5. Flags the listings priced below market, with a plain evidence line ("X% below the median of N comparable sales") and a data-strength indicator so you can judge each number.

Input

Field	Type	Default	Description
manufacturer	string	(none)	Limit the scan to one manufacturer, Hebrew or English (e.g. Toyota or טויוטה). A single manufacturer builds denser cohorts and finds more deals. Leave empty to scan all makes.
maxItems	integer	60	How many listings to score and return rows for. Caps the output only. No upper cap.
corpusItems	integer	1500	How many comparable Yad2 listings to pull as the market baseline. More comparables make valuations more stable.
minDealPct	integer	8	Flag a listing when its price is at least this percent below the fair market price (the threshold is raised when less data backs the valuation).
maxPlausiblePct	integer	60	Discounts larger than this percent are treated as data errors (typos, parts cars) and never flagged as deals.
onlyDeals	boolean	false	On: return only the flagged deals. Off: return every scored listing with its market value and evidence.
proxyConfig	object	Israeli residential	Proxy configuration. Israeli residential proxy recommended (Yad2 is reached from Israel).

For local testing you can also pass listings (an array of Yad2 car listings) to skip the live pull and value against the pool you provide. Pass a representative pool so the cohorts are meaningful.

Example input

{
  "manufacturer": "טויוטה",
  "maxItems": 60,
  "corpusItems": 1500,
  "minDealPct": 8,
  "maxPlausiblePct": 60,
  "onlyDeals": false,
  "proxyConfig": { "useApifyProxy": true, "apifyProxyGroups": ["RESIDENTIAL"], "apifyProxyCountry": "IL" }
}

Output

Each row is one scored listing.

Field	Type	Description
listingId	string	Yad2 listing id
url	string	Link to the Yad2 listing
manufacturer	string	Manufacturer (Hebrew where available)
model	string	Model (Hebrew where available)
subModel	string	Sub-model / trim from Yad2
year	integer	Year of production
km	integer	Kilometers
hand	integer	Number of previous owners
askingPrice	integer	Asking price on Yad2 (ILS)
fairValue	integer	Fair market price from comparable Yad2 listings, adjusted for mileage and year (ILS)
dealGap	integer	Fair value minus asking price (ILS). Positive means priced below market.
dealPct	integer	Percent below market (positive means underpriced)
isDeal	boolean	True if priced below market by at least the threshold with enough comparables
matched	boolean	Whether the listing was matched to a comparison cohort
evidence	string	Plain-language proof behind the price, e.g. "15% below the median of 11 comparable sales"
dataStrength	string	How much data backs the valuation: strong, moderate, limited, or unrated (describes the evidence base, not doubt about the deal)
cohortTier	string	Comparison tier used (see How it works)
cohortSize	integer	Number of comparable listings used for the valuation
cohortMedianKm	integer	Median mileage of the comparison cohort (km)
priceSpreadPct	integer	Price spread within the cohort (robust deviation as percent of median)
kmAdjustmentPct	integer	Percent the mileage adjustment moved the fair value versus the cohort baseline (0 if not adjusted)
cohortDealerShare	integer	Percent of the cohort listed by dealers
kmAdjusted	boolean	Whether the fair value was adjusted for the listing's mileage
suspectReason	string	Why a listing was not flagged (see below)
scrapedAt	string	When scored (ISO date)

Example output row

{
  "listingId": "abcd1234",
  "url": "https://www.yad2.co.il/vehicles/private-cars/abcd1234",
  "manufacturer": "טויוטה",
  "model": "קורולה",
  "subModel": "1.8 היברידי",
  "year": 2019,
  "km": 78000,
  "hand": 1,
  "askingPrice": 92000,
  "fairValue": 108000,
  "dealGap": 16000,
  "dealPct": 15,
  "isDeal": true,
  "matched": true,
  "evidence": "15% below the median of 11 comparable sales",
  "dataStrength": "moderate",
  "cohortTier": "T1",
  "cohortSize": 11,
  "cohortMedianKm": 81000,
  "priceSpreadPct": 9,
  "kmAdjustmentPct": 1,
  "cohortDealerShare": 64,
  "kmAdjusted": true,
  "suspectReason": null,
  "scrapedAt": "2026-06-16T10:00:00.000Z"
}

How it works

The fair value is computed against the live Yad2 market, never an external price guide.

1. Cohorts. Listings are grouped by manufacturer and model (Hebrew names preferred, with edit-distance folding so spelling variants like ספורטז and ספורטג' land in the same cohort). Sub-model and trim are deliberately not part of the key, so cohorts stay dense; trim variance is absorbed by the robust median below.

2. Comparison tiers. For each listing the actor tries the tightest comparison first and falls back only when there are too few comparables:

   • T1 (high): same model, same year, similar mileage. No mileage curve needed. This is the literal pitch.
   • T2 (high): same model, same year, mileage adjusted.
   • T3 (medium): within one year, year and mileage adjusted.
   • T4 (low): any year, same model, year and mileage adjusted.

3. Robust math. Within a cohort the actor uses the median (not the average) and trims outliers with a median-absolute-deviation filter, so one mislabeled, salvage, or typo listing cannot drag the fair value. The mileage effect is estimated robustly from the cohort itself (a Theil-Sen slope), clamped to a sane depreciation band, and falls back to a conservative prior when the cohort is too thin to estimate it.

4. Evidence, shown plainly. Every row carries an evidence line ("X% below the median of N comparable sales") so you can judge each price on the proof, not a label. A dataStrength field (strong, moderate, limited, unrated) says how much data backs it: thin cohorts, missing mileage, fallback mileage estimates, and wide price spreads all lower it. The deal threshold is raised when the data is weaker, so the actor never flags a deal off a thin or noisy cohort. Used-car prices vary widely within the same make, model, and year (mileage, trim, condition), so most rows read "limited" or "moderate" by design, that is honesty about the data, and the evidence line lets the number speak for itself.

When a listing is not flagged, suspectReason explains why:

• subject_data_invalid: the listing's own price, year, or mileage is out of range.
• salvage_suspect: the listing reads as a salvage or parts car (valued, but never flagged as a deal).
• insufficient_comps: not enough comparable listings to value it.
• implausible_gap: the discount is larger than maxPlausiblePct, treated as a data error.
• below_threshold: priced below market, but not by enough given how much data backs it.

Notes and limits

• The fair value is based on Yad2 asking prices of comparable cars, not confirmed sold prices. It reflects what similar cars are listed for right now, which is the relevant benchmark for spotting an underpriced listing.
• Deals only flag when there are enough comparable listings. A rare model with few comparables on the market will be valued at limited data strength or not at all, by design.
• Scanning a single manufacturer produces denser cohorts and more deals than scanning all makes with the same pool size.
• Salvage and parts cars are detected from the listing text and excluded from cohorts; a salvage listing is still valued but never flagged as a deal.
• Prices and mileage are in the units Yad2 uses (ILS, kilometers).

Use cases

• Car dealers and flippers scanning for stock priced below the market.
• Buyers hunting for a specific make or model below its going rate.
• Building a price-tracking dataset of where the used-car market is actually pricing each model right now.