Ridge regression on 1,628 priced jobs across five restoration companies. Enter a scope; get the price range comparable jobs actually came in at.
Both ranges are read straight off held-out prediction errors, not from a standard deviation. Errors on this data are fat-tailed in log space — normality is rejected at p≈10⁻²⁴ — so a normal-theory band would misstate coverage. The inner band (×0.74 to ×1.38) is where half of comparable jobs landed; the outer (×0.52 to ×1.81) holds eight in ten. Both are multiplicative, so they widen with the estimate rather than being a flat dollar cushion — which matches the data: residual spread is near-constant across every size quintile.
Narrowing the band does not raise confidence, it lowers it. Half a standard deviation spans only 44% of outcomes — a range the true price falls outside more often than inside. One standard deviation covers 74% here, more than the 68% a normal distribution implies, because the fat tails concentrate mass near the middle.
Median error is 30%, and 69% of jobs fall within 50% of their prediction. That is useful for triage, sanity-checking an estimator, and spotting an outlier before it goes out. It is not accurate enough to quote a customer from.
Containment and asbestos are the two questions worth asking on the call. Together they cut the residual spread by about 2%, which moves a $7,000 estimate's range by a few hundred dollars — real, but it will not change a decision. Answering "unknown" is honest and supported: the model carries a separate missing indicator rather than reading a blank as "no".
Trained only on jobs whose paperwork carried a room breakdown, and on totals above $1,000 with deposits, draws and equipment-only invoices removed — those were fragments of jobs rather than job prices.
Trained only on jobs whose paperwork carried a room breakdown. Mitigation invoices billed time-and-materials have no rooms to learn from, and the model is materially worse on them — that population is where the residual error concentrates.