Feature importance of char_beds and char_rooms in the Cook County Assessor's residential model
This reproduces the Cook County Assessor's Office (CCAO) residential automated valuation model (ccao-data/model-res-avm) locally to measure how much the number of bedrooms and total rooms actually drive assessed fair market value.
| Item | Value |
|---|---|
| Model | CCAO residential AVM (LightGBM gradient-boosted trees) |
| Data year | 2026 assessment (public inputs from CCAO's S3 bucket) |
| Training rows | ~400K arms-length residential sales |
| Predictors | 95 features (structural, location, proximity, census) |
| Hyperparameters | CCAO defaults (cross-validation disabled; does not affect the importance ranking) |
| Metric | LightGBM built-in gain importance (each feature's share of total loss reduction across all tree splits) |
| Rank | Feature | Share of total gain |
|---|---|---|
| 1 | meta_nbhd_code (neighborhood) |
15.6% |
| 2 | loc_school_elementary_district_geoid |
12.9% |
| 3 | char_bldg_sf (building square feet) |
10.0% |
| 4 | loc_census_tract_geoid |
9.1% |
| 5 | char_class (property class) |
6.3% |
| 6 | loc_school_secondary_district_geoid |
5.5% |
| 7 | acs5_median_income_per_capita_past_year |
4.1% |
| 8 | char_fbath (full baths) |
3.5% |
| 9 | meta_township_code |
2.0% |
| 10 | char_yrblt (year built) |
2.0% |
| 11 | time_sale_day |
1.7% |
| 12 | char_frpl (fireplaces) |
1.7% |
| 13 | acs5_percent_education_high_school |
1.5% |
| 14 | acs5_median_income_household_past_year |
1.5% |
| 15 | prox_num_foreclosure_per_1000_pin_past_5_years |
1.3% |
| 16 | loc_latitude |
1.3% |
| 17 | char_land_sf (land square feet) |
1.2% |
| 18 | char_rooms (total rooms) |
1.0% |
| 19 | time_sale_year |
0.9% |
| 20 | char_beds (bedrooms) |
0.9% |
| Feature | Rank (of 95) | Gain share | Cover | Frequency |
|---|---|---|---|---|
char_rooms |
18 | 1.00% | 1.43% | 1.63% |
char_beds |
20 | 0.90% | 1.32% | 1.24% |
Bedrooms and rooms are minor predictors of assessed fair market value. Each contributes under 1% of the model's total gain, ranking 18th and 20th out of 95 features.
For comparison:
- Building square footage matters ~10x more than either (10.0% vs ~0.9%).
- Full bathrooms matter ~3.5x more (3.5%).
- Location dominates everything: neighborhood (15.6%), elementary school district (12.9%), and census tract (9.1%) alone account for ~38% of total gain.
The likely reason bedroom and room counts carry so little independent weight: they are highly collinear with building square footage, which the model already has. Once a tree model knows a home's size and location, the marginal information in "how many rooms" is small. This matches CCAO's own published guidance, which ranks importance as location > size > bedrooms/bathrooms, and notes strong diminishing returns ("the value added by a second bedroom is much more than the value added by a twentieth bedroom").
- Model trained with
pipeline/01-train.R; importance read from the trained LightGBM booster vialightgbm::lgb.importance()(the same call CCAO uses inpipeline/04-interpret.R). - The linear baseline model in stage 01 was skipped (memory-heavy, and irrelevant to the LightGBM feature importance).
- Gain = share of total loss reduction attributable to splits on each feature. Cover = share of observations touched by those splits. Frequency = share of all splits that use the feature.