Real estate investor
Machine learning valuation models
Rent and after-repair-value models trained on tens of thousands of transactions, served by API with confidence ranges.
- average rent error (MAPE) on ~17K leases
- 6.2%
- average rent error (MAPE) on ~17K leases
- median rent error: half of predictions within 5%
- 4.7%
- median rent error: half of predictions within 5%
- after-repair-value model on ~57K sales, 19 metros
- R² 0.80
- after-repair-value model on ~57K sales, 19 metros
- average after-repair-value error on held-out sales
- ~8%
- average after-repair-value error on held-out sales
Current ARV
After-repair value used for underwriting
ARV
$309,800
ARV $/sqft
$167
Comp basis sqft
1,855
Latest ML model ARV: $305,400 — typically accurate within ±7.3%, i.e. $283,100 – $327,700
Comparable range: $288,500 – $324,900 across 10 comparable sales
Expected sale price by renovation level
| Renovation level | Sale price | Per sqft | Comps range | How fast it sold |
|---|---|---|---|---|
| Great | $298,200 | $161 | $281,400 – $312,900 | half sold within 6 days · 82 sales |
| Great-Excellent (what we usually build to) | $309,800 | $167 | $292,100 – $326,300 | half sold within 5 days · 84 sales |
| Excellent | $321,500 | $174 | $301,700 – $339,800 | half sold within 5 days · 75 sales |
Based on 465 renovated sales nearby. Each extra square foot adds about 68% of what an average one is worth.
The challenge
Rent and resale estimates varied by analyst and lagged the market. Third-party estimates were too generic for renovated rental homes.
Every acquisition decision depends on these two numbers, so small errors compound into bad deals or missed ones.
What we built
01
Purpose-built training data
Models trained on our own transaction history: roughly 57,000 home sales across 19 metros and 17,000 leases, cleaned and filtered for comparable homes.
02
Location-aware features
Hexagonal spatial indexing captures micro-location; property, school and neighborhood features capture the rest.
03
Ranges, not just a point
Quantile models return a low, median and high estimate, so the team can see how confident the model is and underwrite conservatively.
04
Metro-specific where it helps
Dedicated models for metros where they beat the global model (up to 28% lower error in one market), and smart routing between them. The finding: market coherence beats sample size.
Comparable sales (10 properties)
Subject: 1,855 sqft · 4/2 · 2002 · Garage · Target condition: Excellent
Comp summary: $288,500–$324,900 · median $308,450 · avg distance 1.0 mi · avg match 66%
| # | Photo | Match | Address | Sold price | $/sqft | Sqft | Bd/Ba | Year | Distance | Sold | DOM | Reno |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 1 | 69% | 14 Maple Ct | $318,500 | $171 | 1,862 | 4/2 | 2004 | 0.4 mi | 2026-07 | 6 | Excellent | |
| 2 | 67% | 208 Cedar Ridge Rd | $296,000 | $162 | 1,827 | 3/2 | 1998 | 0.7 mi | 2026-06 | 11 | Great | |
| 3 | 66% | 51 Willow Bend | $305,900 | $166 | 1,843 | 4/2 | 2001 | 0.9 mi | 2026-08 | 4 | Great | |
| 4 | 65% | 9 Aspen Way | $324,900 | $175 | 1,857 | 4/2.5 | 2006 | 1.1 mi | 2026-05 | 8 | Excellent | |
| 5 | 64% | 330 Hawthorn Dr | $288,500 | $158 | 1,826 | 3/2 | 1995 | 1.3 mi | 2026-07 | 23 | Great | |
| 6 | 63% | 77 Birch Hollow Ln | $311,000 | $168 | 1,851 | 4/2 | 2003 | 1.6 mi | 2026-06 | 9 | Great | |
| 7 | 62% | 415 Sycamore Ln | $302,500 | $165 | 1,834 | 3/2 | 1999 | 1.8 mi | 2026-04 | 14 | Great | |
| 8 | 61% | 26 Linden Pl | $319,900 | $172 | 1,860 | 4/2 | 2005 | 2.1 mi | 2026-08 | 7 | Excellent |
What we learned
Model accuracy matters less than trust. Confidence ranges and an override path got the team using the numbers every day.
Stack
- Python
- LightGBM
- XGBoost
- CatBoost
- Docker
- Google Cloud Run
Next case study
Field inspection app with AI scopingWhat would you build with one more great operator on the team?
A 30-minute call. We'll ask where the time goes and tell you honestly whether AI can help.