What predicts whether a homeowner redeems
The headline redemption rate answers how often. This answers for whom, and the strongest signal turns out to be the one an auction buyer can calculate before bidding.
Every figure here carries its sample size. Several rest on fewer than fifty windows, which is enough to be a finding and not enough to be precise.
Equity is the strongest observed signal#
The winning bid at the sheriff's sale, as a share of the property's assessed value:
| Winning bid vs assessed value | Redeemed | Windows |
|---|---|---|
| Under 50% | 58.1% | 31 |
| 50–80% | 44.2% | 77 |
| 80% or more | 20.0% | 50 |
Monotonic across all three bands, and a difference of nearly forty points between the extremes.
The mechanism is not mysterious. A lender bidding under half of assessed value leaves a large equity cushion. The owner has something worth saving and routes to save it — refinancing, borrowing from family, or selling the property during the redemption window and keeping the surplus. An owner whose debt approaches the property's value has nothing to redeem for.
The practical inversion: the properties that look like the best deals at auction are the ones most likely to be redeemed out from under the buyer. More than half of them were, in this sample. The certificate holder receives their money back with statutory interest, which is a return but not the property they bid on.
Why Hennepin only. The winning bid amount is published in the Hennepin sheriff's payload and not in the equivalent feeds from Dakota or Washington. Properties with no published bid are excluded rather than pooled into a fourth band, because a missing-data category on a chart appears as though the absence of information predicted something.
Homestead status separates the outcomes#
| Homestead status | Redeemed | Windows |
|---|---|---|
| Homesteaded — owner occupies | 38.6% | 171 |
| Not homesteaded | 24.6% | 114 |
A fourteen-point difference, on larger samples than the bid-to-value cut.
Homestead status is published on the county assessor record, so this is knowable before a sale at no cost. It is the single most useful field on that record for anyone assessing redemption risk.
Normalising it required work. Four county vocabularies had to be reconciled:
Dakota records FULL HOMESTEAD / NON HOMESTEAD / DISABLED VET HOMESTEAD /
FRACTIONAL, Ramsey uses Y / N / P, Hennepin and Washington use Yes /
No. Fractional and partial homesteads are held as a third category rather
than folded into either side, because a fractional homestead is a genuinely
different situation.
One coverage caveat. Anoka records a null rather than N for
not-homesteaded — enumerated across all 140,279 Anoka parcels: Y on 105,887,
null on 34,392, and no third value. A null elsewhere may mean unknown, so Anoka
contributes fewer homestead rows than it has parcels.
Where the rate table and the model disagree#
This is unresolved and it is published because it is unresolved.
A Cox proportional hazards model was fitted on the same data, with county stratification.
| Variable | In the rate table | In the model |
|---|---|---|
| Bid to assessed value | Strongest cut, 20.0% to 58.1% | Contributed essentially nothing |
| Homestead status | 38.6% vs 24.6% | Not significant at conventional thresholds |
| Log amount owed | — | Weakly significant, p ≈ 0.03–0.04 |
Model concordance: roughly 0.66 to 0.70. Better than chance, not strong enough to rank individual windows usefully.
Two explanations are available and we cannot currently distinguish them.
Collinearity. Bid-to-value may be correlated with something already captured — county, amount owed, property type — so the model attributes the effect elsewhere and the marginal rate shows it in full.
A between-county effect. The gradient may exist between counties rather than within them. If low-bid properties are concentrated in counties with higher redemption rates for other reasons, the pooled rate table shows a strong gradient and a county-stratified model sees none.
Why publish the rate table anyway. The rate is an observation — it says what happened in the windows we tracked, with the counts attached. The model is a fit, and a fit that disagrees with the observation is a reason to report both rather than to pick the one that reads better.
A reader relying on the bid-to-value gradient should know the two methods disagree. That is the honest position and it is stated at every point the figures appear.
The homestead artefact, corrected#
An earlier version of the model showed homestead status as strongly significant, and it was an artefact.
Before separating the statutory tracks, 335 tax-forfeiture windows were pooled with mortgage redemption windows. Forfeiture rows are overwhelmingly non-homesteaded parcels, and none of them could ever experience a mortgage foreclosure outcome.
So "homesteaded" was partly predicting "is this a mortgage foreclosure at all", which is not a finding about homeowners.
| Before correction | After | |
|---|---|---|
| Windows | 1,671 | 1,336 |
| Model concordance | 0.803 / 0.805 | 0.661 / 0.669 |
| Homestead-only concordance | 0.742 / 0.761 | 0.594 / 0.618 |
| Homestead p-value | 0.000 | 0.121 / 0.331 |
The disagreement did not resolve — it evaporated. The apparent strength was coming from the pooling error, and once the error was fixed the effect was much weaker. That is a different thing from finding the effect was real, and it is why the raw rate difference is reported without a causal claim attached.
A testable prediction#
Stating one in advance is how this gets checked rather than argued.
Dakota County has an observation imbalance: 52.7% of its homesteaded windows are still pending against 29.3% of non-homesteaded. If homestead status is a real effect, working Dakota's 49 pending homesteaded windows down to resolution should move the county's homestead gap in a predictable direction.
If it does not, the raw difference is largely a coverage artefact — homesteaded properties being checked less completely — rather than a behavioural finding.
We will publish the result either way, with the prediction stated above as the benchmark.
What is not predictive#
Reporting the nulls matters as much as reporting the findings.
Amount owed on its own. Weakly significant in the model and not a usable signal in isolation. A large debt on a valuable property is a different situation from the same debt on a modest one, which is why the ratio works and the raw number does not.
Property type, in the categories available. Single-family versus other showed no useful separation, though the non-single-family sample is small.
Time of year. No seasonal pattern that survives the sample size.
Missing-data indicators. An early model version found these strongly predictive, which was the first sign something was wrong — they were acting as county proxies, because different counties leave different fields blank. Removed.
Redemption rates by county#
Published with the caution that county differences here partly reflect local conditions and partly reflect how completely each county has been checked.
| County | Redeemed | Windows | Confirmed only |
|---|---|---|---|
| Hennepin | 39.2% | 158 | 43.0% of 142 |
| Dakota | 26.2% | 61 | 27.1% of 59 |
| Washington | 24.2% | 66 | 25.0% of 64 |
Hennepin is materially higher, and there are at least three explanations that the data cannot currently distinguish.
Higher property values. More equity in the average distressed property means more owners with something worth saving, which is the same mechanism as the bid-to-value gradient operating at county level.
More complete checking. Hennepin has the richest data feed and the most thoroughly worked backlog. A county checked more completely surfaces more resolved outcomes, and resolutions that turn out to be redemptions raise the rate.
Genuine differences in the local market — employment, housing costs, the availability of family capital, the mix of homesteaded property.
Which is why counties below twenty confirmed windows are not published. A county-level rate on a dozen windows is noise presented as a finding, and it would be quoted as though it were not.
Interpreting a rate table honestly#
Some notes on reading these figures, because the misreadings change the conclusion.
These are observed rates in a sample, not probabilities for a property. A 58.1% figure on 31 windows means 18 of 31 tracked windows in that band ended in redemption. It does not mean a specific property has a 58% chance.
The bands are not equally reliable. 31 windows supports a much wider confidence interval than 171 does. A difference of a few points between two small bands is not a difference.
Correlation is not the mechanism. Equity plausibly causes redemption, and the data shows association. The model disagreement described above is a direct reason to hold the causal claim loosely.
Selection into the sample matters. These are properties that reached a sheriff's sale in counties we track, resolved, and were detected. Each of those filters could bias the sample in ways not fully characterised.
And the confirmed-only column exists for a reason. Where inferred and confirmed figures differ, the gap is the size of the judgement being applied. Hennepin at 39.2% inclusive and 43.0% confirmed is a four-point judgement.
What a buyer should take from this#
Calculate the ratio before bidding. Opening bid from the sale notice, assessed value from the county assessor. Both free, both available in advance.
Check homestead status on the same record. It is one field and it separates the outcomes by fourteen points.
Price the deep discount as risk rather than as opportunity. A certificate at under half of assessed value was redeemed more than half the time in this sample. The statutory interest is real; the property is not guaranteed.
Do not treat any of it as a forecast for one property. The rate describes a population of thirty-one windows.
And remember what the alternative outcome is. A redemption is not a loss — capital returns with statutory interest after a wait. It is a different investment from the one that was intended, and the question is whether that return, over a period that runs past eighteen months in half of cases, is acceptable.
Method and limitations#
Source. Sheriff sale notices from county feeds and legal notices, joined to a parcel spine from county assessor and GIS records, tracked to resolution through recorded deeds and county ownership records.
Classification. Confirmed where a recorded document establishes the outcome directly; inferred where it rests on a judgement. Of 326 resolved windows, 41 are inferred. Headline figures are published both ways.
Sample floors. No cut published below twenty confirmed windows, and none below fifteen for bid bands or buyer types.
Limitations. Minnesota only. Coverage varies by county, which affects county-level comparisons more than statewide ones. The bid-to-value cut is Hennepin only. Samples of 31 and 50 are real findings and not precise ones.
Citation#
Cite freely, with the sample size. A rate without its count is not a finding.
Suggested form: Govire, observed Minnesota redemption outcomes by equity position, n=158 Hennepin windows with published bid amounts, August 2026.