Field Report: Five Weeks of Data From a Self-Learning eBay Pricing Engine

Abstract

Automated pricing for secondhand goods lives or dies on comparable sold listings, and sold-comp search is noisy in ways that are hard to fix with static rules. We built a learning loop that treats every seller correction as a labeled example, mines those examples nightly for repeated patterns, and promotes only the patterns that clear a strict evidence bar. Over the first five weeks the system ingested 137,794 seller events and 194,795 sold orders, ran 29 consecutive nightly learning passes without a miss, and reduced the average share of non-comparable sold comps from 72.9% to 59.5% — an 18% relative improvement in comp cleanliness, achieved without any human labeling effort.

The problem: sold comps are full of impostors

When software prices a used item, the honest way to do it is the way an experienced reseller does: find what the same thing actually sold for recently, and anchor to that. The catch is that querying sold listings is adversarially messy. A search for a specific fuel pump returns lots of pumps that fit different engines. A search for a jersey returns the kids’ size, the signed version, and a four-item bundle. Condition acronyms, seller-specific SKU codes, and filler adjectives all poison the query without announcing themselves.

We measure that mess directly. Before any comp influences a price, a verification layer inspects each candidate and rejects the ones that are not genuine matches. The share it has to reject — we call it the discard rate — is our cleanliness gauge for the whole search pipeline. When the discard rate is high, the searches are wasteful and the pricing evidence is thin. When it falls, pricing quality rises upstream of any model change.

At launch, the number was humbling: roughly 73 of every 100 candidate comps were impostors. That was the baseline the learning system had to beat.

System design, in plain terms

The design premise is that our sellers already know the answers. Every day, resellers on FlowLister generate listings from photos, then fix whatever the AI got wrong before publishing: a reworded title, a corrected item specific, a nudged price. Each of those fixes is a tiny labeled lesson from a domain expert who does not think of themselves as labeling anything. The loop has four stages:

1. Capture. Seller corrections, price revisions, publishes, and sold outcomes are recorded as structured events. In the first 29 days that produced 108,499 listing-edit events, 14,996 price revisions, and 11,943 published listings, joined by 194,795 historical sold orders synced from connected eBay accounts.
2. Mine. A nightly process classifies the full diff history (36,901 individual before-and-after diffs in the most recent run) and asks a narrow question: what do sellers keep correcting, in the same direction, across different accounts? Repeated patterns become candidate rules scoped to a category and field.
3. Gate. Candidates do not go live on a single observation. Each needs repeated, consistent support before promotion, and the bar is deliberately strict: of 17,952 candidate rules mined to date, fewer than 2% have been promoted to production. We would rather learn slowly than learn something wrong.
4. Verify. The comp verification layer stays in place as an independent check. It never learns from the same signal it grades, which is what lets us use its discard rate as an honest scoreboard for everything upstream.

What kinds of things does it learn? The publishable examples are mundane, which is rather the point. It learned that condition acronyms like “NWT” belong in item specifics, not in comp searches. It learned that vague adjectives such as “modern” match nothing useful. It learned that part-number-looking strings sometimes identify an item precisely and sometimes match nothing at all, and in which categories each tends to be true. No single rule is impressive. A few hundred of them, each earned from repeated seller behavior, compound.

Results

The headline series is the weekly average discard rate: the share of candidate sold comps rejected as non-comparable at verification time. Lower is better.

Five weeks, five declines, no reversals. The absolute drop is 13.4 points; the relative improvement is 18%. We find the shape of the curve more persuasive than any single number. Week-over- week monotonic improvement is what a working feedback loop looks like, and the slope steepened in weeks four and five as the promotion gate began clearing accumulated candidates — the compounding phase we designed for.

Reliability matters as much as the trend. The nightly learning pass has completed 29 times in 29 days. A learning system that skips nights does not just pause; it drifts out of sync with the sellers it learns from.

Method notes and limitations

A few honest caveats, because a field report without them is an advertisement. First, this is observational data from a live product, not a controlled experiment; engineering improvements shipped during the window (several of them prompted by what the mining surfaced) contribute to the trend alongside the learned rules themselves. We consider that a feature of the loop rather than a confound to eliminate, but it means the 18% belongs to the whole system, not to any single component. Second, five weeks is a short window and week five is partial; the curve could flatten. Third, everything here is eBay-specific. We make no claims about other marketplaces yet.

What we can say with confidence: the signal source (real seller corrections at scale) is renewable, the evidence bar for promotion is strict, the scoreboard is independent of the thing it scores, and the score has moved in the right direction every week since launch.

Why we are publishing this

Most pricing tools describe their intelligence in adjectives. We think resellers deserve numbers, including the unflattering ones like a 72.9% starting discard rate. Publishing the metric also commits us publicly to keep improving it: this page will be updated as the series extends, and the methodology will stay the same so the numbers remain comparable.

If you sell on eBay and want your corrections to be part of this loop (and your pricing to benefit from everyone else’s), that is what FlowLister does: photos in, a priced and researched draft out, and a system behind it that gets a little harder to fool every night.

Citing this report

Data are drawn from FlowLister production telemetry, June 24 to July 22, 2026. Figures are counts of recorded events, not estimates. To cite:

Frequently asked questions