01 / DECISION · EXECUTIVE SUMMARY
Require local evidence before trusting the price interval.
A real-estate analytics leader needs more than a plausible price. The estimate must communicate where it is dependable and where it needs further review. On later Cook County sales, the flexible spatial model’s small accuracy gain over a regularized property regression is inconclusive.
The local result changes the decision.
In the historical Hyde Park township cohort, those nominal 90% bands cover only 70.1% of sales. That assessment township is larger than the familiar neighborhood of the same name. A countywide average cannot substitute for a local reliability check.
Small gains do not justify false precision.
The spatial-versus-hedonic error difference is -0.21 percentage points, with a 95% interval of -0.53 to +0.18. The result does not establish a clear model upgrade. This is a historical research comparison, not a current appraisal or an individual property quote.
02 / IMPLICATION
Location helps prediction; it does not guarantee local coverage.
Removing location from the boosted model raises median percentage error to 26.48%. Geographic information is useful, but a simpler model with regularized location effects remains competitive. Both principal models tend to underpredict later sales, with median predicted-to-observed price ratios around 0.93.
Use the simpler model as a serious deployment comparator. Assess error and interval coverage by geography and property type, and route unusual or poorly supported cases for further assessment. This study has not tested an operational review process or measured savings.
03 / EVIDENCE
Compare future sale records with a fixed, earlier snapshot.
The final cohort has 24,551 sale records, 23,898 distinct parcels and 331 geographic grid cells. Repeated sales remain clustered in uncertainty estimates.
| Model | Median percentage error | 95% interval | Log-price RMSE |
|---|---|---|---|
| Local training-sale median | 22.63% | 21.59–23.88% | 0.417 |
| Regularized hedonic regression | 17.50% | 16.69–18.49% | 0.340 |
| Spatial gradient boosting | 17.29% | 16.42–18.36% | 0.337 |
| Physical attributes only | 26.48% | 25.23–27.68% | 0.488 |
Median absolute dollar error is about $61,000 for both principal models. Spatial boosting’s countywide median 90% prediction-band width is about $384,244. That wide range is a material part of the result.
HISTORICAL PROPERTY EXPLORER
Does the uncertainty hold in this local market?
Select a historical township and building-size group. Every number comes from evaluated sales; this is not an address-level appraisal or a current market quote.
23,898 distinct parcels · 20,846 earlier training sales in the same township/size group. These are cohort comparables, not a manually adjudicated set for an individual home.
Empirical coverage is shown without a future guarantee. This selection changes the evidence shown, not the fitted model or calibration.
Historical prices and the model’s uncertainty
| Observed sale-price median | $359,900 |
|---|---|
| Observed 10th–90th percentile range | $147,500–$850,000 |
| Median predicted price | $338,189 |
| Median individual prediction-band endpoints | $196,829–$581,073 |
At least 30 final sales are required for each displayed group. Global calibration uses 5,828 earlier parcels. Wider nominal intervals change the saved interval endpoints and measured coverage; they do not improve the model’s point accuracy. No personal sale records or exact property locations are displayed.
Geographic transfer and previously unseen parcels
| Test | Model | Sales | Median error | 90% coverage |
|---|---|---|---|---|
| Unseen spatial cells | Regularized hedonic regression | 5,064 | 19.24% | 86.6% |
| Unseen spatial cells | Spatial gradient boosting | 5,064 | 18.68% | 88.9% |
| Previously unseen parcels | Local training-sale median | 22,938 | 22.16% | 89.7% |
| Previously unseen parcels | Regularized hedonic regression | 22,938 | 16.69% | 90.4% |
| Previously unseen parcels | Spatial gradient boosting | 22,938 | 16.46% | 90.8% |
| Previously unseen parcels | Physical attributes only | 22,938 | 26.47% | 89.9% |
The spatial test excludes 60 final grid cells from fitting, tuning and calibration. The parcel test excludes every final parcel previously present in fitting or calibration. Their populations differ from the primary cohort, so their scores are not interchangeable.
Lowest observed township coverage and the older assessment benchmark
| Township | Sales | Median error | 90% coverage |
|---|---|---|---|
| Hyde Park | 1,309 | 35.7% | 70.1% |
| Calumet | 84 | 34.7% | 71.4% |
| West Chicago | 773 | 26.4% | 76.6% |
| Lake | 2,879 | 27.4% | 78.6% |
| Thornton | 1,313 | 24.0% | 84.0% |
These are exploratory diagnostics, not separately confirmed hypotheses. The frozen prior-year board assessment ×10 is available for 24,498 final sales; 53 lack a positive value. Its median percentage error is 35.09% on that available-case subset. Older tax assessments are not current appraisals or ground truth, and this comparison is not a review of the Assessor’s present operations.
500 paired .03-degree spatial-grid cluster bootstrap draws retain parcel repeats. Split-calibration log-residual bands use finite-sample order statistics; temporal/spatial dependence means empirical coverage is reported, not guaranteed.
04 / DATA
Resolve the property grain before interpreting the price.
The source snapshot contains 1,098,988 improvement/card rows. Restricting it to single-family, single-card, single-land-line and non-prorated properties produces 878,069 unique parcels. Sales are joined many-to-one to those parcels, so a price is never replicated across building cards.
The public sales extract contains 132,773 records from May 2024 through December 2025. Multi-parcel transfers and publisher deed/duplicate/low-price flags exclude 34,637. Another 45,918 do not match the eligible property snapshot; fixed price and area bounds remove 953 more. The analysis covers prices of $50,000–$3 million, buildings of 500–10,000 square feet and land of 500–200,000 square feet.
The characteristics file was last modified April 10, 2024 and describes 2023 records. It predates every included sale. Development uses May–October 2024; November–December supplies validation. Final fitting uses 20,846 May–December sales. January–March 2025 supplies 5,828 calibration parcels, followed by the untouched April–December final cohort.
Sale records were retrieved in September 2026 and may include corrections. Historical ingestion timestamps are unavailable: pre-sale features are verified, but this is not a reconstructed real-time reporting system. Buyer and seller names were excluded from the sales request. No addresses, parcel identifiers or exact property coordinates are published.
Sources: Cook County Assessor’s published model inputs and Parcel Sales. The County terms disclaim accuracy/completeness warranties and endorsement. No separate Creative Commons license is asserted.
05 / METHOD & LIMITATIONS
Keep calibration separate and test geographic transfer.
The local comparator uses earlier neighborhood sale medians with township and county fallbacks. The hedonic model estimates log prices from physical characteristics and regularized location categories. Spatial gradient boosting adds flexible relationships using coordinates and township; validation selects 31 leaves from the frozen 15-versus-31 comparison.
Training-only transformations impute numeric values and encode categories. Later assessments, closing dates, current household characteristics and post-snapshot property changes are excluded. A physical-only ablation measures what changes when location is removed.
Prediction intervals use absolute log residuals from a separate calibration period and a finite-sample order statistic. The 80%, 90% and 95% settings are fixed before final evaluation. Geographic dependence and market drift mean those nominal levels are not guaranteed for a particular property or future market.
- A retrospective sale-date split cannot reconstruct historical sale-ingestion availability.
- Properties with multiple cards, land lines, parcel bundles or prorations are excluded; stated price/area bounds restrict generalization.
- The snapshot becomes stale as properties and markets change; later renovations are not observed.
- Calibration guarantees require exchangeability that geographic dependence and market drift can violate.
- Group median interval endpoints are summaries of individual bands, not confidence intervals for group market value.
- No causal effect, current address-level appraisal, tax appeal recommendation or realized commercial saving is claimed.
- Independent technical review is pending.
06 / CODE
Follow the result back to the frozen inputs.
Run S43-a92f1ea2-98e410fd
Analysis commit a92f1ea2f8e5f77ebb10775722577f8f98162307
git clone https://github.com/mpgibb/michael-gibb-research.git
cd michael-gibb-research
uv sync --frozen
uv run python -W error studies/S43/study.py
uv run python scripts/report_s43.py
uv run python -W error -m unittest discover -s tests -vFixed seeds, locked dependencies and hashes for the snapshot and 20 monthly sale extracts. Mutable API responses must match the frozen checksums. Tests cover parcel grain, timing, interval order statistics, aggregation, suppression and independent reconstruction of primary uncertainty. Independent technical review is pending.