Original research · 370,997 completed lots

Mining data from ~400k auctions for fun and profit

By Ben · September 2026 · Data from 2026-05-06 to 2026-09-01

GovAuctions processes mountains of civic data, much of which is public, interesting, and almost always published in a format that makes it difficult and tedious to parse.

Over the last weekend, I rifled through over 370,000 completed auctions, to see if I could find any fun and unintuitive patterns. Six held up to scrutiny.

The data

Everything comes from the govauctions.app completed-auction archive: 370,997 lots that closed between 2026-05-06 and 2026-09-01 across the US, UK, Canada and Australia. Most of the analysis runs on the 84,134 US lots from 5,323 government sellers where the winning bid is confirmed rather than inferred.

At GovAuctions we have our own model to work out whether a listing is “cheap” or a good deal for a potential buyer. We use sold comps, 3rd party resale stats, location, and several other signals. For this piece I rebuilt a stripped-down version of it that predicts a closing price anchored on listing title, so that every “cheaper than expected” means cheaper than other lots of the same thing. The full methodology is at the end.

So, curious to learn about the strange and sometimes unintuitive world of bidding on government auctions?

1. Some sellers are far less likely to get >1 bid on their listings, making them more profitable to bid on.

If you are bidding: Bid the opening price the day a lot appears from one of the sellers below, and you will usually own it. Those lots close well under what the title alone predicts, though not under our public sold-price comps, which is a distinction worth understanding before you treat it as free money.

Across the whole archive 14.9% of lots sell to a single bidder at the opening price. That rate is stable per seller, so it works as a property of the seller rather than the lot. Most sellers sit near 12.3%. A few sit past 50%.

SY Limited is a Pennsylvania remarketer selling graded coins, silver certificates and trading cards. Six lots in ten go to the only person who bid, at the price the seller asked. Across all 475 sellers with at least 30 closes the pattern is consistent: the quarter with the highest single-bidder rate close 11% below what the title predicts, while the quarter with the lowest close 27.7% above. The correlation across sellers is -0.33.

2. Short listing titles with minimal descriptions sell below what they should, because far fewer bidders ever find them.

If you are bidding: Work a category by hand and open every listing with a two-word title and one line of description. They draw about half the bidders of a fully written listing from the same seller, close 12.6% lower, and a quarter go to a single bidder. The photos are the only thing telling you what the item is.

Nobody searching for a conference table finds a lot called “Table”, no saved alert fires on it, and it surfaces only for people working through a whole category by hand. Lots with a title of two words or fewer and a description of 83 characters or fewer close 12.6% below the same seller's other lots (95% CI -17.6% to -7.2%), on 17.2% fewer bids (95% CI -22.5% to -12.2%) and a median of 7 bids against 13. Against our public sold-price comps they close 12.9% low (95% CI -23% to -1%), on the 376 of them the comp engine can price.

3. Schools and colleges sell items far cheaper than utilities and public works.

If you are bidding: Filter for school and college sellers, where lots close 24% below expectation on a median of 7 bids. Avoid water, transit and public works lots unless you know the equipment better than the contractors bidding on it: those close 24.1 to 42.9% above on 23 to 27 bids. And ignore anything promising police auction bargains.

Police and sheriff sellers show no discount. Comparing vehicles of the same make, model and year, a police seller's car closes -6.3% against everyone else's, on an interval running from -20.8% to +10.1%. The “seized police auction bargain” genre, which is most of what gets written about this market, is not in the data.

4. Sellers who list constantly and at higher volumes are less focused on their margin, so their goods go cheaper.

If you are bidding: Put the ten biggest state, fleet and school surplus operations on a watchlist and sort by seller rather than by ending soonest. The same item costs about 8.9% less from them than from a county selling its only truck this year, which will draw a crowd.

Comparing lots of the same item type, sellers who listed 20 lots or fewer close 8.9% above sellers who listed 500 or more (95% CI +4.9% to +14%). Drop the item-type restriction and the gap is 31.3%, though most of that is what small sellers happen to list. A county that sells two trucks a year has no regular audience, so its one truck is found by everyone browsing at once, at 21 bids a lot. A state fleet pushes supply at the same repeat buyers every week, at 9.

5. A missing physical key takes 38.9% off the closing price of an otherwise identical car, because no bidder can confirm it runs. There is no similar discount for a dead battery.

If you are bidding: Search descriptions, not just titles, for “no key”, “keys not available”, “mileage unknown” and “lien”. Getting a new key costs a fraction of the discount. Skip anything advertising a fresh battery: it pulls 24.4% more bidders and closes 19.7% above the same peers.

Take vehicles and compare lots of the same make, model and year. A dead battery and a missing key are both small, bounded, fixable problems, and the market treats one as an inconvenience and the other as a catastrophe.

The same split runs through the whole corpus at the level of individual phrases. “Mileage unknown” costs almost as much as “not running”, though one is a missing number and the other is a broken engine. The phrase most buyers read as a bargain signal, “as-is, where-is”, sits on 17,913 lots and does nothing at all.

6. The same heavy vehicle sells about a third cheaper in the Northeast.

If you are bidding: Price a haulier before you bid. The gap runs to 34% for the same make, model and year, and PennDOT's near-identical Mack dump trucks repeatedly close on single-digit bid counts. Against public comps the same split shows as 11% below in the Northeast versus +4.5% elsewhere. Shippable categories show no equivalent geography, so this is haulage arbitrage rather than a general one.

Comparing vehicles of the same make, model and year, Northeast lots close 34% below the rest of the country (95% CI -42% to -20.8%). New York is the extreme at 28.5% below. The same vehicle in Illinois, Oregon or Iowa closes 30 to 50% above the national middle.

Rust is the explanation to reach for, and Northeast lots are in worse shape: 25% carry severe damage against 14% elsewhere. But throwing out every lot with severe damage, every non-runner and every impound still leaves a gap of 26.9% (95% CI -35.8% to -15.2%), and the gap is the same size in each of the three months.

Three patterns I wanted to be real, but weren't

Also tested and dead: $1 opening bids (-47.3% raw, -1.2% once you compare within the same seller and item type), pickup-only listings, rural discounts, same-day supply gluts, ALL-CAPS titles and relists. None survived a within-seller comparison.

Method

Expected price comes from a ridge regression on hashed title unigrams and bigrams plus category, platform and vehicle year, fit per country with five-fold out-of-fold predictions, out-of-fold R² of 0.70 and a residual standard deviation of 1.23 log points. Individual lots are unpredictable at that spread, so every figure here is a shift in an average, not a rule about a single auction.

• Comps. Where a figure is quoted against comps it is the public sold-price median the site shows on that lot's own page, recomputed from the current snapshot, so the article and the site never disagree. Real estate and miscellaneous are never priced, which is why some rows read n/a. 44,124 of the 84,134 lots get one.
• Percentages are differences in closing price against the stated comparison group, converted from log points. Intervals are 95% bootstrap confidence intervals. Where one crosses zero I say so and treat the pattern as unproven.
• Controls. Every pattern was recomputed within seller (subtracting each seller's mean residual) and within item type (the first two significant title tokens). For vehicles I used a stricter cell of exact make, model and year with at least four lots.
• Inference. 400 to 600 bootstrap resamples over sellers, not lots. Naive lot-level intervals are far too narrow here and would have kept several of the failures alive.

Data covers 2026-05-06 to 2026-09-01. Prices are winning bids excluding buyer premium and taxes. A pattern that holds across ten thousand lots will still lose you money on any given Tuesday.