How the price outlook works

In one paragraph

We do not know what any single car will be worth. What we can see is today’s asking prices for the same model at every age, and how the asking prices of cars still for sale moved in recent weeks. The outlook turns those two observations into a typical path with an honest range around it. It is a guide for planning a purchase or a sale. It is not a valuation and not a forecast of the market.

Older cars as the future of today’s car

For each model, in each country with enough data, and for all eight countries pooled:

1. We take every car of that model on sale today whose registration year is published (at least 5 cars from 3 dealers in that year, no dealer holding more than 60%).
2. We fit one curve through their asking prices by age: the logarithm of the price against age, allowed to bend at 2, 4, 7, 10 and 15 years (where the data reach that far), with an adjustment for fuel (and for country in the pooled curve), so that a year full of diesels does not look cheaper only because of its fuel mix.
3. The expected change for a car of a given registration year after 6, 12, 24 or 36 months is the step along that curve from its age today to its age then. We never go past the oldest cars we see: if there are no cars three years older, there is no 36-month figure.

What this means in practice: “if today’s pattern held, a 2021 Golf in Germany would be listed about 10% lower a year from now” says that, in today’s market, Golfs one year older are listed about 10% lower, after allowing for fuel. It is a scenario: it assumes the market stays where it is today, and older cars also differ in mileage, trim, condition and sometimes generation.

The range (80%) combines two things:

• how much the curve itself moves when we refit it on a different mix of dealers (the dealers drawn again at random, with replacement, 100 times);
• how far individual registration years sit off the curve. Model changes, new generations, a facelift or an unusual year move whole years away from a smooth line, and both your car’s cohort and the older cohort that stands in for its future can be one of those (a car keeps its registration year as it ages). This part is estimated for each model and never set lower than the average of all models: no model is assumed to be steadier than the typical one.

When the top of the range is more than 25% above its bottom, the figure is not shown: it would say too little.

While a car is for sale

The second part follows individual adverts. For each car still advertised one, two or four weeks later, we compare its asking price then with its asking price at the start. We publish one average change per country, with all models combined, because separate figures per model predicted the next week no better in testing. To set the indicative range, we measure how far each model group’s average change was from its country’s figure in the latest week held out for checking. We pool these distances across all eight countries and use their 95% split-conformal quantile to set the same width around each country’s figure. Coverage within each country and in later weeks has not been verified. See the table below for current results.

• Only cars that stayed on sale are counted. A car that sold, or was withdrawn, leaves the sample: we never see sale prices.
• A price that jumps by more than about 25% in a few weeks, or by exactly a VAT rate (19% to 23%), is set aside as a likely data error (for example, a switch between prices with and without VAT).
• A source only counts from the week its recorded price changes reached a steady rate. We have no log of each re-read, so this is a proxy. Until late October 2026 the large marketplaces recorded few price changes, so the first figures rest mostly on dealers’ own websites.

Price history: weekly snapshots

Every Saturday at 00:00 UTC, we snapshot every used-car listing on sale in the eight countries at its latest asking price, judging the price valid or not as it stood that week. Duplicates within one source are removed. Whether a listing counts is fixed the first time we record it, so a later change of its status cannot rewrite earlier weeks: a listing marked sold, or switched to a price without VAT, later counts until the night we first see that; one first seen already marked sold (about 0.5%) is left out, because we do not know when it sold; Standvirtual adverts listed net of VAT (about 0.3% of that source) are left out, because their earlier prices carry no record of that basis. The weeks before 3 November 2026 were reconstructed on that day from the listings’ state then, and are marked as such. For each model and country, and all eight together, we keep counts of listings and dealers and the 25th, 50th and 75th price percentiles. We compute a week nine days later, after re-reading cars still listed. The newest two weeks are provisional and recomputed nightly; older weeks are frozen. History starts on 3 October 2026. The partial crawl week of 26 September is excluded.

A point needs at least 30 listings from at least five dealers, with no dealer holding more than 60%. The chart needs eight such weeks. Cross-portal de-duplication covers only cars on sale today, so a car on two portals counts twice each week. The listing mix follows crawl coverage. Many marketplace listings left our crawl in the week of 19 October 2026. The weekly median shows what was for sale, not a price trend.

The like-for-like line uses cars on sale in consecutive weeks from sources whose price changes we were already recording in the first week (inferred from each source’s rate of recorded price changes, not from a log of every re-read). We average changes in log asking price, set aside jumps over about 25% or exactly a VAT rate as likely errors, and chain the result into an index starting at 100. Sold or withdrawn cars leave; ageing and remaining unsold still affect the line. Until 26 October 2026, major marketplaces’ price changes were not recorded, so early weeks use fewer cars and mostly first-seen prices.

Market range: Eurostat

We use Eurostat’s monthly Harmonised Index of Consumer Prices for “second-hand motor cars”: dataset prc_hicp_minr, COICOP 2018 code CP07112, index 2025 = 100. It is available by country and for the EU (27 countries). From each monthly starting point since January 2015, or the first month with country data, we measure the index change three, six and twelve months later. For each horizon, the range is the size of change that at least 80% of those past periods stayed within, centred on no change.

We centre it on no change because carrying the previous twelve months’ trend forward was not reliably better at twelve months. The table below compares the average errors of both approaches by country. This is a national consumer price index for all used cars, with methods that differ by country. It does not measure this model’s asking price. The measured periods overlap, so they are not independent. They also include the 2021–22 shortage shock.

Read the market range beside the age-effect range. Do not add them. Past index moves do not guarantee future ones, and coverage of future outcomes has not been validated. Source: Eurostat. Reuse requires attribution; the ranges are our calculation.

Frozen outlooks for later scoring

Once a week, we save a copy of the published outlook. It is taken from the files of the first successful publication of the week, with that build’s identifier and time, and contains the age-effect figures, the median asking price and count used as their starting point, and the market ranges. Later builds in the same week are not saved. The age figures will be scored against the published median of the same registration year at the target date (a stock-price test that includes changes in the mix of cars, not an isolated age effect); the market ranges against the index of the base month plus the horizon, as first published. Keeping that copy lets us assess the outlook as it appeared at the time, using the figures readers actually saw. A later change to the current outlook does not change the saved copy.

When three, six and twelve months have passed, we can compare each saved outlook with what happened. We plan to report bias, coverage and range width by country and by horizon. Bias will show the direction of average error; coverage will show how often outcomes fell within the published ranges; width will show how wide those ranges were. Those measures need later outcomes before they can be reported.

The first three-month comparisons become possible in about February 2027. Until then, the label remains: “Model-based; forward coverage unvalidated.”

Limits

• Asking prices, not sale prices. Dealers often accept less than the advertised price.
• Different cars. Older cars today are not your car in the future: they differ in trim, mileage, condition and sometimes generation.
• Short history. Weekly price histories started on 26 September 2026. The value curve has not yet been tested against real months passing; that test becomes possible as the history grows.
• The market moves. Taxes, fuel prices, a new model or a change in interest rates can move the whole market. The outlook assumes today’s market.

How we tested it

What the first tests showed (data of 2 November 2026).

• While for sale: the first release’s per-model 95% range (mean ± 1.96 standard errors) contained the held-out result only about 84% of the time at one week and about 74% at two weeks: it was too narrow. The table below compares the errors from per-model and country-wide figures with “no change” and gives the current range and pooled coverage. The indicative range now uses split-conformal calibration on the latest held-out week, with one shared width set from model groups across all eight countries. Coverage within each country and in later weeks has not been verified. Set the same way one week earlier, it contained only about 63% of the next week’s results, because marketplace re-pricing began to be captured in between. Read it as the latest spread, not a guarantee; we will re-check it as weeks accumulate. The two-week horizon has no earlier week available yet to test its range in this way.
• Value curve: on dealers held out of the fit, the curve’s one-year steps missed by about 6 percentage points on average. The error the size of the held-out sample alone would cause is estimated at about 5 points (an estimate that assumes independent cars). The raw step between two registration years did slightly better on that test; the curve is kept because it gives the 6, 24 and 36-month figures and smooths over thin years. The 80% range, widened by that estimated noise, contained the held-out comparison about 80% of the time; it measures comparisons between registration years, not the future price of one car.

The tables below are recomputed with every nightly update.

Weekly points used: 26 Sept 2026 to 24 Oct 2026. Groups shown: 1,354 value curves, 9 for-sale changes.

While for sale: predictions compared with observed changes

The first release's per-model range was too narrow. The indicative range now uses a shared width set from model groups across all eight countries in the latest week held out for checking. Set the same way one week earlier, it contained only about 63% of the next week's results, because marketplace re-pricing began to be captured in between. Coverage within each country and in later weeks is unverified. Read it as the latest spread, not a guarantee; see the table for current figures.

Value curve: dealers held out of the fit (one in five)

Comparisons (one-year steps between consecutive registration years)	1,148
Average error of the one-year value change	6.11 pp
Error expected from the held-out sample's own size alone	4.83 pp
Average error of the raw year-to-year step of the other dealers	6.04 pp
Inside the 80% range (widened by the held-out sample's noise)	81%

Market range: Eurostat's used-car price index

Every monthly start from 2015-01 (or a country's first month) to the latest month with data (2026-09). Centred on no change, the range holds at least 80% of the past changes. The two average errors compare “no change” with carrying forward the past 12 months' trend, on the same starting months.

Source: Eurostat, HICP – second-hand motor cars (prc_hicp_minr, index 2025 = 100). Ranges and errors are our calculation. Eurostat dataset