The Cook County Assessor values residential property with a model whose source code, training data, feature list, and performance statistics are public. Most assessors will not show you their model. This one publishes it under an open license and documents every change.
The documentation for the 2026 assessment year includes this note: after the initial run, the Assessor "was informed by the Illinois Department of Revenue that some relevant home sales from 2025 were missing from our database. After ingesting these new sales, the model was re-run, so there are *two* final 2026 models."
- **February 11, 2026:** townships valued - Berwyn, Bremen, Cicero, Lyons, Oak Park, Palos, River Forest, Riverside, Stickney; sales in training set - 396,480.
- **July 30, 2026:** townships valued - Bloom, Calumet, Lemont, Orland, Proviso, Rich, Thornton, Worth; sales in training set - 405,195.
The Assessor adds that "the two models perform almost identically" and that the office judged "even such a minor improvement was worth re-running the model."
Quick Answer
For assessment year 2026 the Assessor published two final residential models after IDOR flagged missing 2025 sales: a February 11 run (396,480 training sales) covering nine townships and a July 30 run (405,195 sales) covering eight. The office says the models perform almost identically. Record which run valued the parcel-especially for uniformity arguments that cross the township split-and use the public feature list as proof of what the model cannot see.
What this does and does not support
It does not support an argument that the first nine townships were valued on a defective model. The Assessor's own statement is that the difference was minor, and a firm claiming otherwise without its own ratio study is doing the thing the Board's rules call cherry-picking, in reverse.
It supports two narrower points. The Assessor treats completeness of the sales file as material enough to re-run a county-wide model, which is a useful frame when a firm argues a specific sale was missing or mis-recorded. And a parcel in Berwyn and a parcel in Proviso were valued by different runs on different training sets, which matters if a uniformity argument crosses that township line. Record the run in the file. It costs nothing, and it matters if the file reaches PTAB with an appraiser who wants to know what the county did.
The feature list is a list of what the model can see
The residential model uses roughly one hundred inputs: bedrooms, square footage, year built, basement type, roof material, and a large block of location factors (school district, neighborhood code, distances to transit, water, and parks, walkability, flood risk). It trains on residential sales from the nine years before the assessment date.
Read the list as an inventory. Anything not on it, the model cannot have priced: interior condition, deferred maintenance, a failed foundation, the neighbor's junkyard. That is the whole basis of a condition argument, and the Assessor has published the proof.
What the Board actually grants on
The Board of Review's decision history records the reason for every reduction. For attorney-filed appeals in tax years 2023 through 2025, the most common grant reason, 190,274 decisions, is "analysis of comparable properties, a recent sale, and/or an update of property characteristics."
An update of property characteristics. The Board's own top grant code names the case where the county's record of the house was wrong. Square footage, basement type, class. The public feature list tells a paralegal exactly which fields drove the value and therefore which errors are worth finding.
Comparables: the Assessor's own words
The repository is careful here. The Assessor does not use sale or uniformity comparables "for the purpose of modeling." The model does generate comparables through an experimental feature, and the documentation says they "are experimental and are not currently used" for valuation, though they are used for appeals and for evaluating model performance.
So when a firm submits a comparable grid, it is not correcting the Assessor's comps. It is offering evidence in a form the Assessor did not use, and the argument should say so: the model cannot see condition, and here are five sales of houses in comparable condition it could not distinguish from the subject.
The performance statistics, used honestly
The Assessor evaluates each model on a held-out test set and publishes COD (coefficient of dispersion), PRD (price-related differential), PRB (price-related bias), and MKI (modified Kakwani index), the same measures the assessment profession uses to judge uniformity. They describe the model across the county. They support a general statement that mass models have measurable dispersion. They do not, alone, show that one parcel sits at the wrong end of it. The parcel-level argument still needs the parcel-level evidence.
What a paralegal does with this
- Pull the parcel's characteristics from the Assessor's open data and check each against the feature list. An error in the county's record is an error in the model's input.
- Note which of the two 2026 runs valued the township.
- Build the condition record the model cannot see: dated interior photographs, contractor estimates, inspection reports. The feature list is the proof these were never considered.
- Cite the county-wide statistics as context, never as the parcel's proof.
What not to do
Do not call the model a black box. It is the opposite, and the Board's analysts know it. Do not call it biased in a township without running the ratio study; the Assessor publishes the tools to run one. Do not treat the experimental comparables as an admission. The documentation says they are not used for valuation, and it says so plainly.
The strongest use of a public model is the narrowest: here is everything the county considered, here is what it could not consider, and here is the evidence for the part it missed.
The numbers behind this article
Model quotations and the two-run table are from the Assessor's public model repository as of September 5, 2026. Decision counts are Censum's tabulation of the Board of Review Appeal Decision History dataset (county open data, 7pny-nedm), copy dated May 20, 2026; PIN-level decisions, tax year 2025 incomplete. Censum is independent and not affiliated with the Assessor or any county office. Censum's parcel record shows the county's characteristic fields next to the evidence a firm gathers, so the gap between what the model saw and what the house is fits on one page.
Sources
- Cook County Assessor model-res-avm repository
- Cook County Board of Review Appeal Decision History (7pny-nedm)
- Censum tabulation of the Cook County Board of Review Appeal Decision History dataset (Cook County open data, 7pny-nedm), copy dated May 20, 2026; PIN-level decisions; tax year 2025 incomplete in that copy.
FAQ
Which townships were valued on which 2026 run?
February 11, 2026: Berwyn, Bremen, Cicero, Lyons, Oak Park, Palos, River Forest, Riverside, Stickney (396,480 sales). July 30, 2026: Bloom, Calumet, Lemont, Orland, Proviso, Rich, Thornton, Worth (405,195 sales).
Does the two-run fact prove the first nine townships were defective?
No. The Assessor states the difference was minor. The useful points are that sales-file completeness mattered enough to re-run, and that parcels in different township sets were valued on different training sets.
What is the Board's most common grant reason on attorney-filed appeals?
For tax years 2023 through 2025: 190,274 decisions coded to analysis of comparable properties, a recent sale, and/or an update of property characteristics.
Does Censum file or represent clients?
No. Censum LLC is independent property-tax intelligence and docket tooling. It is not a law firm. The attorney files and makes every legal judgment.
Next step
Look up each PIN's township against the two-run map, pull the county characteristic fields, and build the condition record the model cannot price. The attorney decides how to frame the gap.
Censum LLC is independent and not affiliated with Cook County or any government agency.