UQO Working Paper No. 36· v1.1· November 2026· 69 pages
Price Discovery, Inventory Liquidity, and Dealer Repricing in Quebec's Used-Car Listings
A Pre-Registered Descriptive Study of a Daily Panel of 273,265 Dealer Listings, August–November 2026
One used-car listing in five is marked down (median 2.8 percent) and the markdown hazard doubles over three weeks, but only 3 of six pre-registered tests reject their null after Holm adjustment.
A daily panel of 273,265 Quebec dealer listings, read about seven times a day over 81 days. The nonlinear age–price gradient and the rising markdown hazard hold; the competitor response is not supported, regional premia are not detectable at the precision of the data, and the liquidity index does not pass its internal predictive check. Asking prices are not transaction prices and a listing's disappearance is not a verified sale.
Simon-Pierre Boucher — contact@spboucher.ai
19.7 %
Share of listings that receive at least one sustained markdown; the median first cut is 2.8 percent of the entry price and the daily markdown hazard doubles over the first three weeks before plateauing. Listings entering above their make–model mean or their local competitors are cut sooner and more deeply.
+2.8 %
Improvement in dealer-grouped out-of-sample fit from age and mileage splines over a linear specification, which is rejected. Age is identified across model-year cohorts over an 81-day window: a cross-cohort gradient, not within-vehicle depreciation.
0.29 vs 0.35 points
Same-day association between a close competitor's price cut and an own cut, once the same dealer's other feeds are purged, against the placebo of pseudo-competitors more than 200 km away. It does not exceed the placebo, and the lagged response sums to −1.00 points: the pre-registered competitor response is not supported.
±3.7 points
Range of the 15 quality-adjusted regional premia relative to Montréal. The joint test does not reject with trim fixed effects (p = 0.486) and rejects with make–model fixed effects (p = 0.006); dealer-clustered intervals cannot exclude premia of up to 9.9 points in 4 remote regions.
26.6 %
Share of confirmed exits lasting under 1.5 days, non-market events by the structure of the feeds. On all exits the relative-price coefficient is a bounded null; without these spells, listings priced above their make–model mean exit more slowly (hazard ratio per 10 percent between 0.924 and 0.969).
94 % / 10.2 %
Matched exits that coincide within a week across sources, and exits followed by a relisting within 30 days: exit is a consistent event that overstates retail sales by at least a tenth.
p = 0.076
Internal predictive check of the pre-registered five-component liquidity index on the 16 named regions (slope 0.34 percentage points per unit, wild-cluster bootstrap): it does not pass at the 5 percent level, and no external validation is possible with an 81-day window.
Abstract
Asking prices and listing exits are the only vehicle-level, high-frequency signals on Quebec’s used-car market. This pre-registered descriptive study follows 273,265 dealer listings observed by the aggregator Auto-Ka from 12 August to 1 November 2026, each asking-price change and disappearance recorded about seven times a day. Hedonic models reject a linear age–price gradient and improve dealer-held-out fit by 2.8 percent. One listing in five receives a sustained markdown (median 2.8 percent); the markdown hazard doubles over three weeks and is higher for listings priced above their make–model mean or their local competitors. Quality-adjusted regional premia lie within 3.7 percentage points of Montréal, but dealer-clustered intervals cannot exclude premia of up to 9.9 points in 4 remote regions. Without the 26.6 percent of exits lasting under 1.5 days, listings priced above their make–model mean exit more slowly (hazard ratio per 10 percent between 0.924 and 0.969). Same-day co-movement of competitor price cuts does not exceed a region-swap placebo once same-dealer feeds are purged, and the pre-registered lagged response is absent. A five-component liquidity index does not pass its internal predictive check at the 5 percent level on 16 regions. Of six pre-registered tests, 3 reject their null after Holm adjustment. Asking prices are not transaction prices and exit is not sale; the paper documents what listing data can and cannot say.
JEL: D22, D83, L11, L62, R32, C41
Key Results
Daily asking-price paths of the dealer segment
The sequence of asking prices of each listing, the decision to cut and the moment of disappearance are documented at the daily frequency for Quebec dealers, where earlier listing studies were cross-sections.
A nonlinear, powertrain-specific age–price gradient
Splines in age and mileage reject linearity and improve held-out fit by 2.8 percent; younger battery-electric cohorts are priced lower relative to their model-year cohort than comparable combustion vehicles (age slopes −0.094 against −0.085).
Markdown dynamics on the lot
Cuts are frequent but shallow and increase with listing age; listings priced above their make–model mean or their local competitors are cut sooner and more deeply, a result that holds with trim-level and dealer-demeaned residuals. The number of local competitors is not stably associated with the markdown hazard.
Co-movement is synchronisation, not competitive reaction
27.1 percent of the first version's competitor pairs were the same physical dealer on another source. Once they are purged, the same-day co-movement does not exceed a region-swap placebo, the lagged response is negative and the intraday ordering is symmetric; no causal claim is made.
What listing exit does and does not measure
Exits are decomposed by likely cause and checked across sources and against relistings of the same vehicle identification number: a quarter are feed artefacts, at least a tenth are relistings, and the true share of retail sales among exits is unknown.
A transparent liquidity index with documented limits
A five-component index summarises 42 percent of the variation in regional turnover but does not pass its pre-registered internal check; deep markdowns turn out to be a sign of slow, not liquid, markets.
Quality control
This paper went through a multi-stage quality control: 7 quality-control reports and 90 points checked. The version presented on this page is the revised manuscript. Scores are on a 0–100 scale and are a diagnostic of the manuscript, not an acceptance probability.
Score after revision
71.3/100
64.0 before quality control · +7.3
Quality control dated November 7, 2026
- Quality-control decision
- major revision
- Scientific integrity
- PASS
- Reproducibility
- R1reproduced from documented analysis data
- Maturity
- Asubmission-ready
- Quality-control reports
- 7
- Points checked
- 90
Scores by dimension
Before quality controlAfter revision
- Originality & contribution67.6 → 67.0
- Scientific rigor61.9 → 74.0
- Identification & methodology62.0 → 70.0
- Data & reproducibility62.9 → 72.0
- Robustness & validation66.7 → 74.0
- Writing70.9 → 72.0
- Figures & tables62.4 → 75.0
- Literature positioning59.6 → 70.0
Revised manuscript69 pages11 figures19 tables3 equations141 references30,195 words
What quality control changed
Quality control changed three conclusions, and the revised paper states them as such: the competitor response is no longer supported, the relation between relative price and exit became a bounded null on all exits, and the liquidity index fails its internal predictive check. The nonlinear age–price gradient and the markdown results stand; the title was changed and the study is now presented as pre-registered and descriptive.
- Competitor co-movement: a same-day co-movement of 0.9 percentage points among close competitors → 0.29 points once same-dealer feeds are purged (27.1 percent of the first version's pairs were the same dealer), not above the 0.35 points of the region-swap placebo; the pre-registered response is not supported.
- Liquidity index: 56 percent of common variation and a predictive next-week relation → 42 percent of the variation in regional turnover and an internal check not passed at the 5 percent level (p = 0.076) on 16 named regions; the first version's larger slope rested on a pseudo-market of listings with a missing region.
- Regional premia: within four percent of Montréal and jointly insignificant with trim fixed (p = 0.53) → within ±3.7 points (p = 0.486), with intervals that cannot exclude premia of up to 9.9 points in 4 remote regions; equivalence within ±5 points for 10 of 15 regions.
- Exit and price: exit described as strongly predicted by price → on all exits the relative-price coefficient is a bounded null; negative only once the 26.6 percent of exits lasting under 1.5 days are removed.
- Depreciation: nonlinear depreciation with a three percent gain in held-out fit → a cross-cohort age–price gradient (not within-vehicle depreciation) with a 2.8 percent gain.
- Scorecard added: of six pre-registered tests, 3 reject after Holm adjustment, one of them in the direction opposite to the one registered.
Limitations stated in the paper
- Every price is a dealer's posted price: negotiated discounts, financing and trade-in terms are unobserved, and the asking–transaction gap is unknown.
- Listing disappearance is an event on a website; the true share of retail sales among exits is unknown, and the exit results describe inventory turnover on dealer websites, as bounds for spells under 81 days.
- Eleven weeks of late summer and autumn cannot identify seasonal patterns; the regional non-rejection applies to one quarter and is bounded only loosely.
- The competitor event study is descriptive and cannot separate a response from common shocks; the liquidity index has no external validation and an internal check it does not pass.
Data
Auto-Ka daily listing panel
273,265 distinct listings observed over 81 days (12 August to 1 November 2026, 586 synchronisation cycles) by an aggregator built and operated by the author, which re-reads 131 dealer websites, two full-inventory portals and five dealer groups, giving 7.3 observations per listing per day and logging every asking-price change and disappearance.
Analysis samples
A hedonic sample of 136,393 listings from reliable sources and a dynamic sample of 80,856 listings that also excludes flickering connectors, oscillating listings and the 47,235 listings already present on the first crawl of their source. Two sampled portals and private sellers are observed but excluded; every excluded listing is written to an exclusion log.
Cross-source exits and relistings
Vehicles observed simultaneously on a dealer's own site and on a portal, and relistings of the same vehicle identification number elsewhere, are used to check exit as a sale proxy. An exit requires seven days of continuous absence.
Regions
16 administrative regions observed, with Montréal as reference; region effects are identified from dealers located in different regions (20 to 175 per region). The registry is not a random sample of Quebec dealers and no coverage ratio can be stated.
Methodology
Hedonic asking-price models
Log entry price on cubic B-splines in age and mileage, vehicle characteristics, make–model or make–model–trim fixed effects, entry-week, source-class and region effects, with standard errors clustered by dealer and dealer-grouped out-of-sample evaluation.
Markdown and exit hazards
Hazards of a first sustained markdown and of listing exit as functions of listing age, local competition and the entry price relative to the make–model mean, with Cox and accelerated-failure-time variants; exit is re-estimated without the artefact spells and with cause-specific hazards for relisting.
Competitor event study with placebos
A dealer's repricing around the price cuts of competitors within 25 km, with market-by-week and model-year-by-week fixed effects, a competitor set purged of the same dealer's other feeds, a region-swap placebo and region-clustered wild-cluster bootstrap inference; estimates are treated as associations.
Regional tests with bounds
Joint tests of quality-adjusted regional premia with trim and with make–model fixed effects, equivalence tests within ±5 points (established for 10 of 15 regions) and minimum detectable premia by region (3.9–8.8 points).
Liquidity index and internal check
Exit intensity, markdown frequency, markdown depth, quality-adjusted dispersion and arrival rate are standardised and combined; the check regresses next-week exit intensity on an index built without the exit component, with market fixed effects and a wild-cluster bootstrap over 16 market clusters.
Pre-registration and multiple testing
Hypotheses, estimands, samples and tests were written down before estimation; the six primary tests are reported with Holm-adjusted p-values, and secondary tests are unadjusted and labelled exploratory.
Reproducibility
- The code (MIT licence), the results files that generate every exhibit and in-text statistic, and the manuscript are deposited on Zenodo and Figshare, versioned with the manuscript.
- The pipeline carries an immutable snapshot, an exclusion log, the pre-registered plan, a deviations table and a claims ledger, and is designed to be rerun on a longer panel.
- The raw snapshot contains vehicle identification numbers, URLs, free text, images and exact coordinates and is not distributed.
- An anonymised processed layer (hashed identifiers, rounded coordinates) is regenerated from the snapshot in under a minute, but its public deposit remains an open governance decision; until then third parties can reproduce every exhibit from the committed results files but not from the raw data.
Cite this paper
Boucher, S.-P. (2026). Price Discovery, Inventory Liquidity, and Dealer Repricing in Quebec's Used-Car Listings. UQO Working Paper No. 36, Université du Québec en Outaouais.
@techreport{boucher2026wp36,
author = {Simon-Pierre Boucher},
title = {Price Discovery, Inventory Liquidity, and Dealer Repricing in Quebec's Used-Car Listings},
institution = {Université du Québec en Outaouais, Département des sciences administratives},
number = {36},
type = {UQO Working Paper},
year = {2026},
month = {nov},
url = {https://www.spboucher.ai/research/wp36},
}More in Finance & Economics