Three fill rates, one dataset, and the cost hiding between them

The operation

A single fulfilment centre serving two banners inside a six-node network that also ships from a second FC and four stores. Twenty-four months, August 2024 to July 2026.

The thesis

The supply-chain discipline has not settled this. Fill rate resolves at least three ways, and published guidance states plainly that there is no agreed OTIF formula because the calculation depends on which point of view you are measuring and the level at which the data is stored. This module is not inventing a strawman; it is naming a documented condition of the field and then resolving it for one operation.

The certified metric register

Generated from the meta blocks on the dbt models that compute these metrics — the same file that defines their tests. Nothing here is typed twice, and build_metric_register.py --check fails the build if a definition drifts from the model behind it.

Certification is not a ranking. An exploratory metric is not a bad metric; it is one the business has agreed not to run on. Retaining the other two definitions and labelling them is the governance act — deleting them would move the disagreement rather than resolve it.

The pathology this replaced

Before certification there were three fill rates in circulation and no statement of which one the business ran on. Each was defensible to the team that built it: Operations measured what physically moved, merchandising measured assortment coverage, Finance measured whether the customer got what they asked for.

The cost was not licences or dashboards. It was that the three teams could not have the same conversation about the same week. Operations reported service improving while Finance reported it flat, both were arithmetically correct, the meeting resolved nothing — and the split rate, which none of the three measured, went unexamined for the entire period.

What the data shows

Descriptive, then diagnostic, then prescriptive. Every chart carries its own data table and its own provenance.

Descriptive — the three rates never meet

The gap is 7.0 points between the most forgiving definition and the strictest. It is not measurement error and it is not a reconciliation problem: each rate is a different question asked of the same shipped quantities. Order fill is arithmetically the lowest of the three in every period, because an order counts as filled only if every line in it is.

Diagnostic — splitting is how the network achieves fill

For every order, the simulation also records what the same order against the same inventory position would have shipped if splitting were forbidden — the best any single node could have done alone, evaluated before the real allocation was committed. Splitting buys 2.65 points of unit fill. Without that counterfactual, “splitting is how the network achieves fill” is an assertion; with it, the governance decision has a price.

Shortfall and splitting are not two problems. They are two symptoms of one condition: inventory fragmented relative to what customers actually put in a basket. Slow-moving styles are ranged at two nodes rather than six, so a basket containing one either reaches for a second node or goes short — and which of those happens is a matter of luck, not of policy.

The second finding — the same rule, two different economics

The dollar cost of a split barely differs between the banners. Its consequence differs by more than three times, because it is charged against a margin that differs by more than three times. A split that is a rounding error on a premium order is a material loss on an off-price one.

There is therefore no single correct split threshold. That is a governance finding, not a modelling inconvenience: any single network-wide rule is simultaneously too aggressive for one banner and too permissive for the other.

Prescriptive — priced, with the residual reported as a residual

The recommendation: set the threshold per banner, not per network. Holding Off-Main splits above $6 as exceptions would catch 5,464 orders carrying $58,218 of split premium. Applying the same rule to Alder & Vance would catch 2,426 orders and save $25,048, while risking margin that comfortably absorbs the cost.

The saving is not free and the page will not pretend otherwise. Those orders shipped complete because they were split. Holding them puts 14,944 units at risk of shipping short, which would move the certified metric — order fill — in the wrong direction. The decision is a trade between a measured cost and a measured service loss, and the register now shows both next to each other, which it could not do before.

The residual, stated as a residual. A per-banner threshold at $6 addresses $83,266 of the $248,933 total split premium. $165,667 remains, in splits too small individually to hold but numerous enough to matter in aggregate. Nothing in this analysis explains that portion away, and no allocation rule tested here removes it. It is reported rather than absorbed into the recommendation.

Where the cost actually lands

Is this world plausible?

The operational data is invented, so the question a sceptical reader should ask is whether the generator produced a world a real department store would recognise. Three audits check it against real published data, and each one is written so that it can fail.

The modelled network runs at 0.97 months of stock against a real department-store band of 2.03–3.64. Sitting below the band is the correct answer, not a miss. The Census series covers entire department stores, most of whose inventory sits on selling floors serving walk-in customers this module never simulates. So the audit asks for the relationship rather than the level: the series must sit strictly inside the sector figure, clear a floor below which the modelled service level would be unattainable, and move with the real series seasonally. It correlates at +0.62 and dips every November and December, exactly as department stores do when peak-season sales outrun the inventory behind them.

The second audit bounds this DC’s operating cost at 6.83% of shipped sales against the 23.7%–39.0% SG&A range Macy’s, Kohl’s and Dillard’s actually report. The third requires splitting to be concentrated and directional rather than uniform noise — which is the pattern shown above, and which a generator producing random splits would fail.

An audit that cannot fail is not an audit. Each one is a pure function of measured numbers, so the validation suite hands each deliberately wrong values and confirms it trips: inventory inflated past the sector band, seasonality flattened, cost inflated past peer SG&A, cost cut to a rounding error, splits made uniform, and concentration reversed. All six trip, and that result is recorded next to the real one in the validation report.

Data & method

Honest limits