Accuracy & calibration

Error distribution

n = 400 held-out subjects

Within 5% (PPE5PPE5The share of valuations within 5% of the realized price — a tighter accuracy bucket than PPE10.)

58.0%

Within 10% (PPE10PPE10The share of valuations within 10% of the realized price — a standard automated-valuation accuracy benchmark.)

88.3%

Within 20% (PPE20PPE20The share of valuations within 20% of the realized price — a looser accuracy bucket used to gauge tail behaviour.)

99.5%

Mean abs error (MAPEMean absolute % errorThe average of |predicted − actual| ÷ actual. More sensitive to outliers than the median (MdAPE).)

5.2%

RMSERoot mean squared errorSquare-root of the average squared error — penalizes large misses more heavily than average error. (penalizes large misses)

6.6%

Median CIConfidence intervalThe range (here, 90%) within which the true value is expected to fall. A tighter interval means a more certain estimate. width

26.1%

Confidence-interval calibration

For each nominal interval level, the share of realized prices that actually fell inside. A well-calibrated model sits on the diagonal — our intervals mean what they say.

Calibration is slightly conservative (intervals contain the truth a touch more than the nominal rate) — we prefer that to overconfidence on a credit decision.