WP36 — Research questions (preregistered, 2026-11-01, before any main-model estimation)
Central question. How do dealer competition, vehicle characteristics, inventory age and local market conditions jointly determine asking-price dynamics and the liquidity of a durable good (used cars) in Quebec?
Data regime (from the Phase 0 audit). Listing-level panel of asking prices from Auto-Ka, 2026-08-12 → 2026-11-01 (81 days, ≈ 7 crawls/day), 273 265 listings, 138 direct dealers + 2 full-inventory portals (≈ 900 dealer names), no transaction prices. All outcomes are asking prices and observed listing disappearance.
Confirmatory hypotheses (tested exactly as specified in analysis_plan.md)
| ID | Hypothesis | Primary outcome | Unit | Primary test | Direction / falsifier |
|---|---|---|---|---|---|
| H1 | Conditional asking-price depreciation is nonlinear in vehicle age and mileage and differs by powertrain and segment. | log first sustained asking price | listing | Hedonic model A with tensor-product splines in (age, km); Wald test of nonlinearity (spline vs linear) and of interaction with powertrain (EV/hybrid vs ICE) and body segment | Falsified if linear-in-age-and-km specification is not rejected at α = 0.01 and the out-of-sample RMSE gain of the spline is < 0.5 % |
| H2 | The hazard of a first asking-price cut and the conditional cut size depend on listing age, local competition and vehicle attributes. | first sustained markdown (event), cut size (%) | listing-day | Discrete-time complementary log-log hazard with listing-age dummies/splines; Cox with PH check; cut-size OLS on markdown events | Falsified if the listing-age profile is flat (joint test p > 0.05) and the competition density coefficient has a 95 % CI covering 0 |
| H3 | A price cut by a close local competitor is associated with a higher subsequent repricing hazard of nearby comparable listings, beyond market×time and product×time shocks. Association, not causal. | repricing (cut) indicator in the window after a competitor cut | listing-day | Event-study / distributed-lag hazard with leads and lags, market×week and model-year-cohort×week FE; clustered by local market | Falsified if lead coefficients are as large as lags (no lead–lag asymmetry) or if placebo (random pseudo-competitors) gives similar estimates |
| H4 | Quality-adjusted asking-price differences persist across Quebec administrative regions net of model-trim-year cohort, mileage, condition proxies, dealer type and time. | hedonic residual / region premium | listing | Model E: region effects in the hedonic model with cohort FE; within-specification dispersion | Falsified if region effects are jointly insignificant (p > 0.05) and all |
| H5 | Relative asking price (hedonic residual at entry) predicts listing disappearance; disappearance is an imperfect sale proxy. | observed disappearance (exit) | listing-day | Discrete-time and Cox exit hazards with the entry residual; concordance; cross-source consistency of exit timing (direct site vs otogo) and relisting detection | Falsified if the residual coefficient's 95 % CI covers 0 in the dynamic sample |
| H6 | A transparent market-level liquidity index (AMLI) built from inventory exit intensity, listing-age distribution, markdown frequency/depth, quality-adjusted dispersion and arrival rates captures common variation in observed turnover across regions and segments. | AMLI (daily/weekly, by region & segment) | market-day / market-week | PCA/equal/supervised weights; bootstrap CIs; leave-one-component-out prediction of next-week exit intensity (held-out component) | Falsified if the first component explains < 40 % of component variance or if the leave-exit-out index has no predictive association with next-week exit intensity (CI covers 0). External forecasting value: NOT testable in this window (declared). |
Exploratory analyses (reported as such; no confirmatory claims)
- Depreciation surfaces by make and by model-year cohort; EV vs ICE vs hybrid residual depreciation.
- Repeated markdowns and markdown sequences (size of 2nd/3rd cuts); price increases.
- Dealer-level repricing styles (frequency, size); banner/chain vs independent dealers.
- Private-seller (Kijiji) vs dealer asking prices for the same make-model-year (descriptive only; no VIN).
- Dealer-to-dealer vehicle transfers detected by VIN relisting at another dealer (wholesale channel proxy).
- Prediction models (Model F) with LightGBM; chronological, dealer and region holdouts.
Interpretation constraints (apply to every result)
- Asking price ≠ transaction price; listing disappearance ≠ sale (may be withdrawal, wholesale, relisting, feed error).
- 81-day window: seasonality, macro shocks and long-run dynamics are unidentified; the day-1 stock is left-censored.
- Dealer coverage is an operator-built registry, not a random sample of Quebec dealers.
- H3 is conditional association; H6 forecasting value is untested.