Key takeaways
- Median error tells you how far off a typical estimate is.
- Hit rate tells you how often estimates land within a set distance of the sale price.
- Always test on sales the model hasn't seen.
Accuracy is measured against real sales
The fair test is simple: take homes that sold recently, hide the sale price from the model, ask for an estimate, then compare. Doing this on thousands of sales gives an honest picture.
Median absolute error
If a model has a 4.8% median error, half of its estimates were within 4.8% of the actual sale price, and half were further off. Lower is better.
Hit rates
A hit rate answers: how often was the estimate within 5% (or 10%) of the sale price? A model might hit within 5% on 55% of homes and within 10% on 82%.
Why accuracy changes by market
Dense suburbs with similar homes and frequent sales give models plenty to learn from. Rural, luxury and one-of-a-kind homes give them very little. That's why we report accuracy by market and price band, not just one headline number.
