Used Car Market Analysis — SHAP Report

What does the model look at when it builds a price? The model explained here is LightGBM — the one called "the model" in the technical report. CatBoost's two variants appear only in the §4 comparison. It works at the level of the whole dataset: which feature takes what share, and how its effect changes across the range. At the single-listing scale there is only one example, in §6.

What SHAP is, and how it differs from LOFO

SHAP splits a single prediction into parts: on this listing age pushed the price up this much, mileage pulled it down that much; the parts add up to the prediction exactly. The ablation/LOFO (leave-one-feature-out) in the technical report asks how much the error grows when a feature is removed.

How it was computed

• Data: all 29,988 listings — no sampling.
• Model: OOF. Every listing is explained by the fold model that never saw it in training — the chain's own 5-fold setup. The rebuilt folds' predictions match the chain's stored OOF predictions (largest difference ₺0.00).
• Method: shap.TreeExplainer (exact). Additivity error 5.1e-13 (measured on the final model), so the parts reconstruct the prediction exactly.
• Scale: target log1p(price), so contributions add in log and multiply on price (see the table below).
• Grouping: the 170 SVD dimensions of the model/series name mean nothing individually, so they are summed into MODEL_SERIES. In the §3 table ENGINE = hp + cc and DAMAGE = the 12 features derived from 13 panels plus the heavy-damage record; in the beeswarm and cohort charts they appear as separate rows. The §6 breakdown only merges the SVD dimensions (MODEL_SERIES); every other feature stays separate — there power has its own row; displacement's share on that listing is small, so it stays inside the "other features" bar. A group's value is the signed sum of its members on each row — the group's net effect on that listing; the table reports its mean absolute value (§7).

Which scale in which chart

Each chart uses the scale that is most meaningful to the reader:

Worked example. The mean |SHAP| of Age (years) is 0.2870 → e^0.2870 = 1.332: a contribution of that size moves the price by 33.2% (the exponential of a mean overstates the typical effect; the typical value is in the table). If Engine (hp + cc) adds 0.1347 on the same listing the percentages do not add — the multipliers multiply: 1.332 × 1.144 = 1.525 → 52.5%, not 33.2% + 14.4% = 47.7%.

What sets the price

These shares are specific to LightGBM (§4). Typical effect on price is the median over listings of the contribution's price equivalent (|e^s − 1|).

Age (years) alone takes 43.9% of the attribution, the top three 77.4%.

The ranking is not fully stable across folds: the item in position(s) 3, 4 changes in at least one fold; do not read much into the order of items with close shares.

Direction check

Direction of effect (Spearman rank correlation): age -0.99, mileage -0.96, engine power +0.89. More age and mileage push the price down, more power pushes it up. This is a sanity check enforced in the generator: if any sign flips, the report is not produced.

The three curves:

• Age: median contribution +90.4% at 0–2 years and -47.0% at 18+, a 3.6× gap in the credit the model gives age; the actual median price ratio of the same two groups is 5.7×. The difference comes from old cars also differing in other features; SHAP books that part to them — mostly Mileage, Engine (hp + cc), Damage (panels + heavy damage).
• What mileage costs (difference between consecutive window medians; window centre = the window's median mileage; scaled to 100k km because the windows are not evenly spaced): 8.7% between 28k and 127k km, 11.4% between 127k and 223k, 7.2% between 223k and 385k. The rate drops clearly in the last window (0.64× the previous one): at very high mileage extra kilometres cost less, but they still cost. At 350k km and above there are 1,723 listings.
• Engine power: median contribution across power bands <150 hp -14.9% → 150–200 hp +7.7% → 200–250 hp +15.5% → 250+ hp +43.2%; from sub-150 hp to 250+ hp the gap in credit is 1.7×.

shap's own split: two groups

We did not pick the split: shap's own decision tree cut the data at Age (years) = 9.5. The mean |SHAP| of Age (years) is 0.396 in Age (years) < 9.5 and 0.230 in Age (years) >= 9.5. Reading that as "the model looks at age more on new cars" misleads: |SHAP| measures the deviation from the average, so at a typical age the age contribution is small. Mean |SHAP| by age band: 0–4 0.577 · 5–8 0.333 · 9–12 0.084 · 13–16 0.271 · 17+ 0.611 — V-shaped; age weighs heavily on both new and very old cars.

Age and mileage act together

Colour is the car's age. In the same 150–250k km band the median mileage contribution is -0.030 under 10 years (2,232 listings) and -0.006 at 10 or older (9,000 listings): the model charges less for mileage on an older car. The two features are not independent (§7).

Three variants build the same price for different reasons

On accuracy the three variants are close (MAPE 6.49% · 6.44% · 6.58%). They differ on reasoning: LightGBM gives age 43.9%, CatBoost (SVD) 30.9% (13.0 points apart); on mileage it reverses (12.9% · 22.1%). Age and mileage move together, so which one gets the credit is the model's preference. Takeaway: the top three items are the same in all three models (Age (years), Engine (hp + cc), Mileage), but their order and shares depend on the model; in CatBoost (native) the 3rd and 4th items are only 0.4 points apart.

This table compares the three models in their final form and under the same rule as §3: a group's value is the signed per-row sum. In the native variant the model and series names are two separate features; here they are summed per row as well. The §3 shares come from the OOF models, the LightGBM column here from the final model; the same item differs by at most 1.4 points between the tables (Model/series name).

Why brand is next to nothing

brand has mean |SHAP| 0.0001, 0.0% of the attribution. The model name already determines the brand (technical report §3: U(brand | model) = 1.00); adding brand on top of series+model changes the mean error by ₺1.

Three methods — dependence, ablation, SHAP — land in the same place: brand carries no separate information. That is not "brand does not affect price"; its effect sits inside the model name.

How one prediction is built

At the bottom E[f(X)] is what the model predicts before it sees any feature, at the top f(x) is its prediction for this listing; each arrow between them is one feature's contribution, and the "15 other features" arrow is the rest summed (log scale, §2).

• Listing: 118i Standart · 2007 · 330,000 km
• Actual price: ₺595,000
• Model's prediction (it never saw this listing): ₺519,680 (-12.7%)

The three biggest contributions:

1. Age (years) = 19 yr → ×0.54
2. Model/series name = 118i Standart → ×0.87
3. Engine power (hp) = 138 hp → ×0.89

The other 21 items together multiply the price by 0.76 (-23.5%) — not small in total.

No bar accounts for the 12.7% gap: SHAP decomposes the prediction, not the actual price. What the model does not know is not written in the breakdown.

Limitations

• Attribution, not causation. SHAP says what the model used, not how the market works.
• Features that move together share the credit. How age and mileage split it depends on the model (§4). TreeSHAP runs here in tree_path_dependent mode, i.e. observational, which also shapes the split.
• Contributions live in log space. When comparing items, read the multipliers: they mean the same thing on every listing. Converting to lira is listing-specific and order-dependent, so this report does not do it.
• The grouping is a decision. MODEL_SERIES is the signed per-row sum of 170 SVD dimensions, i.e. the name's net effect on that listing; taken one by one each dimension looks small. Summing the members' |SHAP| separately was not used: that sum grows with how many pieces the input is split into — the same name summed over its 170 dimensions comes out 2.5× larger.