Methodology note — Why we filter outlier sales before aggregating

This is the July 16, 2026 entry in our rotating methodology series — sixteen explainers cycling through the statistical concepts and data assumptions that sit behind every Bagonomics research piece. The series is meant to be read once and referenced often. Today's entry covers why we filter outlier sales before aggregating.

Outlier filtering is one of the most consequential and least visible steps in our aggregation pipeline. Every transaction that enters our cross-platform sample is checked against a statistical fence before it is allowed to contribute to a median or to a sample-size count. The reason is simple: a single bad transaction can move a brand-level median by an amount that no real market move could produce.

The fence we use is the Tukey one-and-a-half-times-IQR rule. For a given variant in a given 90-day window, we compute the twenty-fifth and seventy-fifth percentile of observed sale prices. The interquartile range is the gap between them. The fence is set at one-and-a-half IQR below the twenty-fifth percentile on the low side and one-and-a-half IQR above the seventy-fifth percentile on the high side. Transactions outside the fence are flagged as outliers and excluded from the median computation.

The Tukey rule is the standard outlier filter in the descriptive-statistics literature. It was designed for exactly the situation we face: a distribution with potentially heavy tails where the analyst wants to keep the legitimate sample and discard the obviously bad ones. The choice of one-and-a-half — as opposed to the more aggressive one-and-a-quarter or the looser two — is calibrated to keep about ninety-nine percent of a normal distribution while clearing typing errors that miss by an order of magnitude.

What gets filtered out in practice falls into three categories. The first is data-entry mistakes — a transaction recorded at thirty-two dollars when the actual sale was three thousand two hundred, or at three hundred and twenty thousand when it was thirty-two thousand. These are usually catchable by other means but the filter is the backstop. The second category is condition-misclassification — a heavily-worn variant sold for restoration-project pricing that gets miscoded as standard condition. These are real transactions but they do not belong in the same aggregation as new-or-pristine sales. The third category is fraud-detection escapes — listings that turned out to be problematic but that crossed our pre-aggregation checks anyway.

What the filter does not touch is the legitimate tail of the distribution. A Birkin in a particularly rare color, or with full set including original receipt and box and dust bag and authentication, can legitimately clear forty or fifty percent above the variant median and still be inside the Tukey fence as long as the IQR of the variant is wide enough. We do not want to filter these out. They are part of the true price distribution and they belong in the median calculation.

The interaction between outlier filtering and sample size matters. Filtering reduces the effective sample. For variants with already-thin samples, an aggressive filter could remove enough transactions to push the variant under the high-confidence threshold. We monitor for this and surface the pre-filter and post-filter sample sizes in the aggregation row for transparency.

The Tukey rule is symmetric — it filters equally on the high side and the low side. This matters in markets where the underlying distribution is asymmetric, which luxury-handbag prices typically are. Our distributions are skewed to the right; there are legitimately more high-tail observations than low-tail observations. The Tukey rule removes a higher proportion of high-tail observations than low-tail in absolute terms. We have considered an asymmetric variant of the filter and we will publish data on the side-by-side impact when the methodology revision is finalized.

The headline rule for the reader: every median we publish is a Tukey-filtered median. If a brand-level number looks unexpectedly clean compared to the underlying transaction sample, the filter is doing the work of removing the visible noise.

Where this comes up in our research

This concept anchors a number of Bagonomics analyses. The research archive surfaces the live applications; the methodology hub collects the full set of these notes for sequential reading.

Methodology

Part of the Bagonomics daily editorial rotation — a 14-day cycle of daily research pieces. Each day's slot is selected from the rotation by day-of-year so the same calendar date always lands on the same topic. Data is frozen at publication; live numbers are visible on the linked entity pages. Methodology notes are editorial explainers — they do not contain time-varying computations and the body text is stable across reruns. The slug rotates by date so that each calendar day in the cycle has its own URL, keeping the daily-publication cadence intact.

*Snapshot frozen at publication. Daily editorial rotation — see /research for the full archive. This is statistical analysis, not investment advice.*