The Better Values Flywheel
We don't just hand you a model and walk away. We run a continuous improvement cycle — clean your data, enrich it, model it, validate it, and keep getting better. Here's exactly how it works.
Get a Demo
...or read on to learn more.
STEP 1
Data Validation (Pre-Model) — Stop Bad Data in Its Tracks
Put a thermometer in your data.
We examine 119 data fields across every parcel in your jurisdiction and produce a Data Coverage Report. This tells you, at a glance, what percentage of your data is clean — broken out by property class, neighborhood, and asset type.
For each category, you can drill down to see exactly which fields are missing, which have errors, and get a downloadable CSV with the specific parcels that need attention. No synthesis required — just a clear list of what to fix and where.
Why it matters:
If it's garbage in, it's garbage out. It almost doesn't matter what model you run — regression, cost, income — if the data feeding it is wrong, the values will be wrong. We start here because everything downstream depends on it.
AI Quality Grading
Same standard. Every property. Every time.
We trained an AI model on the Marshall & Swift quality grading rubric. It assigns a quality grade. We also search online listings for interior photos when available.
The output is a report showing every parcel where the computer's grade disagrees with your current grade. Your team reviews and QCs the differences — the computer doesn't override anyone. It just gives you consistency and catches the ones that slipped through.
Why it matters:
Every appraiser has a slightly different idea of what a Q3 is. When five people are grading properties across a jurisdiction, inconsistency creeps in. This normalizes it.
Automated Sales Validation
The computer takes the first pass. Your team makes the call.
We check every sale for red flags: multi-parcel transactions, grantor-grantee matches (related party sales), same-day same-price sales, and statistical outliers within each cluster of similar properties.
Flagged sales appear in a simple UI where your appraisers can confirm or override the computer's recommendation. It's not replacing their judgment — it's saving them from having to manually screen every transaction from scratch.
Why it matters:
Bad sales in your dataset poison your ratios and your model. Catching non-arm's-length transactions early means better values downstream.
STEP 2
Data Enrichment — Better Inputs = Better Outputs
50+ parcel tags you'd never have time to build yourself.
We enrich every parcel with features that matter to value but are painful to collect manually. These are calculated automatically from your GIS shapefiles, federal data layers, OpenStreetMap road networks, and other public sources.
What we tag: Corner lots, street frontage, cul-de-sacs, waterfront, golf course proximity, topography, flood zones, zoning restrictions (height limits, parking requirements, deed restrictions), lot shape and rectangularity, street lighting, adjacent street type, and more.
Why it matters:
These features affect value. If you're not capturing them systematically, you're leaving equity on the table. And you're certainly not going to send staff out to flag 35,000 parcels for corner lot status. We do it in a batch run.
Automated Neighborhood Delineation
Redraw your neighborhoods using math, not memory.
If you're like most offices, you've been adding neighborhoods for decades. Every new subdivision gets a new neighborhood code. Twenty years later, you have hundreds of neighborhoods and nobody remembers why half of them exist.
We use the Louvain algorithm (from network science) in two steps. First, we draw physical boundaries — rivers, highways, ridgelines, major road networks. Then we cluster parcels within those boundaries by quality, condition, effective age, lot size, and price per square foot.
The result: clean land economic areas (LEAs) with consistent pricing tables, and neighborhoods within them that reflect actual market behavior.
Why it matters:
Better neighborhoods = better land pricing tables = more defensible values. And you didn't have to redraw a single line by hand.
STEP 3
Modeling — Smarter, Fairer Values
Nine models. One answer. More accurate than any single approach.
We build nine independent mass appraisal models and take the median estimate — an approach called the "wisdom of crowds."
The analogy: if you phone a friend on Who Wants to Be a Millionaire, they get it right about 30% of the time. If you poll the audience, they get it right about 90% of the time. Nine independent models, each with a slightly different perspective, consistently outperform any single model.
The nine models:
- •Explainable Linear Multiple Regression (MRA) — The classic. Originally developed in the 1920s for farmland valuation. Popular in the Northeast.
- •Geographically Weighted Regression (GWR) — Adapts coefficients to local market conditions across your jurisdiction. Originally developed in forestry. Compute-intensive but highly equitable.
- •XGBoost Gradient-Boosted Trees — Modern machine learning. Extremely accurate, but requires cross-validation and holdout sets to prevent sales chasing.
- •Vacant Land Models — Estimates each parcel as if vacant, providing a defensible land/improvement split.
Every model uses the sales comparison approach. We just use nine different ways to get there — residential, commercial, vacant, and everything in between. The ensemble median is smack on a 1.0 ratio.
STEP 4
Data Validation (Post-Model) — Double-Check for Confidence
The model reveals what the data couldn't.
After the initial model run, we do another full round of data validation. Modeling surfaces data errors that weren't visible before — a property coded as 1,200 sq ft that the model says should be valued like a 2,400 sq ft home, a parcel graded Q4 that's surrounded by Q2s and priced accordingly, a neighborhood where every ratio is 20% high because of a systematic data entry error.
We flag those anomalies, your team investigates and fixes them, and the next model run is even better. This is where the flywheel starts to compound — each pass through the cycle catches things the previous pass missed.
Why it matters:
No model is better than the data behind it. But a good model can tell you where your data is lying to you. This step is what separates a one-and-done revaluation from a system that actually improves over time.
STEP 5
Deliver + Repeat — Stronger Fairness and Equity Every Cycle
This isn't a one-time project. It's continuous improvement.
We deliver land values, improvement values, and total values for every parcel — along with a value range for each property backed by ratio statistics that give you confidence you'll be compliant.
What you get:
- •Land, improvement, and total values for every parcel in your jurisdiction.
- •A value range per property with defensible ratio statistics — so you know you'll hit your COD targets before you certify.
- •Market adjustment factors for your neighborhoods, derived from our statistical models.
- •Diagnostic ratio study statistics — a clear picture of how you're doing, broken out by property class, neighborhood, and price tier.
Then we do it again. The flywheel keeps turning. Each cycle produces better data, which produces better values, which produces fewer appeals and more defensible assessments. Your office gets stronger every year — not just at the revaluation, but between them.
See this with your data.
Get a Demo
We'll run your jurisdiction through the flywheel and show you the results before you commit. Data quality report, enrichment, neighborhoods, and model values — all with your actual parcels.
Get a Demo
Valuebase is on a mission to build the world's first assessor intelligence tools.
