1,213 listings across five sale dates, joined to the city's parcel, crime, permit, licence, violation, tax-lien and deed records — including 5,593 sheriff's deeds since 2021 that show what properties in each neighbourhood actually sell for.
The August sale has closed, so its outcomes are known. They are the single best guide to what the four remaining sales will do — and they contradict the obvious read of an opening bid.
Twenty-one properties were won by third-party bidders in August. The median one sold for six times its opening bid, and the range ran from 1.05× to 14.78×. A $20,000 opening bid is realistically a $120,000 purchase.
Worse for the buyer: of 414 listed, only 79 reached the block at all — the rest were postponed or stayed — and the bank bought back 58 of those. Your odds of taking a given listed property home were about 5%.
Opening bid → winning bid. Sorted by price.
| Address | Open | Won | × |
|---|
Closing is two independent hurdles: the property has to reach the block at all, and then you have to outlast the bank. August gives a measured rate for both — and only one variable predicts the second.
Only 17% of August listings were sold at all; the rest were postponed or stayed. Postponement is a delay rather than a cancellation — 78% of them were re-listed on a later date, and 213 of the 411 candidates here are August carry-overs.
Nothing in the data predicts which listings get postponed: debt-to-value was 0.77 for those that sold and 0.74 for those that didn't. Treat the 17% as a flat toll on every candidate.
Debt-to-value is the one variable that separates a third-party win from a bank buy-back, and it is close to deterministic at the extremes.
| Debt ÷ city value | Won | Bought back | Win rate | Candidates |
|---|
Not one of the twenty properties with debt above 1.2× its value was won by a bidder. Fisher exact against the under-0.5 band: p = 2.8 × 10⁻⁶.
In August the winners sat in visibly better areas — half the violent crime of the buy-backs. That signal reverses in the larger sample. Across 5,593 sheriff's deeds since 2021 the third-party share rises with neighbourhood distress, because banks do not bother reclaiming cheap stock: third-party deeds run a median $35,100 against $67,800 for institutional ones.
| Violent-crime quartile (400 m) | Deeds | Third-party share |
|---|
Two samples, opposite signs — so the August crime gap is a small-sample artefact of low-debt properties happening to sit in better areas. Only debt-to-value survives both. It is the sole closeability input below.
Median third-party sheriff's-deed price within 400 m of the candidates in each ZIP, from deeds recorded since 2021. This is the number to underwrite against — not the opening bid. The third-party rate is the share of nearby sheriff's deeds that went to a buyer rather than an institution; where it is low, the bank takes the house back.
Hover for detail.
| ZIP | n | Clears at | City value | Ratio | 3rd-party | Violent |
|---|
411 residential candidates under $50,000 across the four remaining sales, scored for a buy-and-hold rental. Clears at is a gradient-boosted prediction of the third-party sale price, trained on 3,594 historical deeds; hover it for the p25–p75 range. All-in adds any delinquent city tax. Value is the city's market value divided by all-in — above 2× is where the thesis works.
| # | Sub-scores |
|---|
Gradient-boosted trees over 3,594 third-party sheriff's deeds since 2021, predicting the log of sale price ÷ assessed value. Scored with 5-fold GroupKFold across 250 one-kilometre spatial blocks, so a property is never validated against its own neighbours — random folds would flatter it.
Spatially honest cross-validation. The baseline is a flat multiple of assessed value.
| Estimator | Median error | Within ±25% | Within ±50% |
|---|
A real gain, and a hard ceiling. 30% median error means an auction is not precisely predictable — who shows up on the day dominates. Rank with it; never set a maximum bid from it.
Permutation importance, mean absolute error.
Assessed value carries it, but tax-delinquency density is the second strongest signal — distressed blocks clear at a steeper discount to assessment than clean ones.
Every city record within 400 m of each parcel centroid is counted, converted to a percentile within the 411-property pool, and weighted for a rental hold.