Economics Before Metrics

Start by pricing the error, not by measuring the metric.

What does a miss actually cost? Not in service points. In the thing the organization loses: a stranded repair job, a technician idle for a day, a customer who buys the next one somewhere else, cash tied up that had another use.

The moment those numbers exist, most rankings become obvious. A write-up whose top line is “forecast accuracy will improve” reports a finding about a diagnostic. A write-up whose top line is a replenishment rule, supported by holding cost and missed-job cost, recommends a decision. The house metrics may still appear in the second, as instruments beneath the reasons, never as the reasons themselves.

Price the error before you defend a metric

Ask what a miss costs, ask what an extra unit of stock costs, and ask whether those two costs are symmetric. They almost never are, and that asymmetry is frequently the whole decision.

You do not need a perfect cost model. You need one better than the proxy it replaces, and good enough that changing the cost assumptions would be visible in the ranking. If the ranking is insensitive to a factor of two in the cost of a miss, stop arguing about the third significant figure. If it is sensitive, that sensitivity is not a footnote: it belongs in the evidence as a condition on the recommendation.

Precision that cannot change the pick is complexity that has not earned its cost, presented as accounting.

The practical bar is lower than people fear. At Meridian Motors a miss on a water pump that strands a vehicle is not the same object as a miss on a floor mat. Last year’s traces plus a planner’s account of what actually stops a job are enough to put those in different classes. You do not need an activity-based cost of every item. You need to know whether the costs are of the same order.

That is also the argument for treating items by class. The economics are unlike; the house metrics are alike because they are easy to put in a pack, and ease of reporting is not an objective. The same holds for an expedite rule, which is a threshold on which parcels fly: “improve on-time” contains no ranking, while the freight bill and the cost of a stop contain it.

Metrics that pretend to be constraints

“We need 98 percent service” sounds like a constraint. It is usually a position with a history, and it should become a constraint only once it is clear which interest it stands for. Sometimes that is a contractual penalty that can be priced, which belongs in the economics. Sometimes it is a vice president who was blamed for a shortage once, which belongs in the conversation about interests.

“Inventory must fall 20 percent” is similar. When there is a genuine cash event, a covenant or a quarter-end or a promise already made, the number is close to a constraint for this cycle, and the live choice is how to meet it with the least damage. When the number originated on a slide, price it or send it back.

So a fill target and dock capacity do not belong in the same list of constraints. Dock doors are physical; the fill target is, until shown otherwise, a position. Putting both into a solver turns the ranking into a search for a point that satisfies two positions at once, which often does not exist and is then reported as a surprising infeasibility.

When a technical metric earns its keep

Technical metrics are not banned. They diagnose the engine; economic consequences rank the alternatives.

Use forecast error when debugging a forecast that feeds a named rule, not when scoring the business. Use utilization to find a queue, then model the queue. A model that improves forecast error by three points and changes no order is a diagnostic with a press release.

A diagnostic can appear in the write-up as a reason only if its being false would change the ranking. “Forecast error is 31 percent” does not meet that test. “Even a large improvement in forecast error would not move the items that hold the cash across class boundaries” does, because it is a claim about work deliberately skipped.

Ranking when the units will not reduce

Pricing the error works when consequences can be pushed into a common unit. Sometimes they cannot, and pretending otherwise produces a worse answer than admitting it.

Alternatives routinely differ on cash released, on stranded jobs, on how many hours a month planners spend maintaining classifications, and on how exposed a rule is to dealers going around it. The first two can be forced into dollars with some violence. The last two resist.

The usual response is a weighted score, and it is the worst available option, because it buries the judgment in coefficients nobody can interpret. A weight of 0.3 on planner hours means nothing until it is converted back into the exchange rate it was standing in for.

Do this instead. Lay the alternatives out with one column per consequence, each in whatever unit that consequence naturally has: dollars for cash, jobs for misses, hours for maintenance, a plain sentence for bypass exposure. Normalize nothing yet.

Then take the two eliminations that require no judgment at all. Remove any alternative beaten by another on every column; it cannot win under any exchange rate. Remove any column on which all surviving alternatives score about the same; it cannot discriminate between them, and carrying it forward only adds an occasion for disagreement. Do both out loud.

At Meridian the incumbent falls to the first elimination, because the class-based rule beats it on every column. Planner hours fall to the second, because finance’s blanket cut (every item down 20 percent by the freeze, then the current rule at targets 20 percent lower) and the class rule cost about the same to maintain. What remains is the blanket cut against the class rule, on cash and stranded jobs, and now the trade has to be faced one exchange at a time: what change in one column would exactly compensate for giving up the other?

Put both columns on the same clock first

The two columns are not on one clock. The blanket cut releases about 400 thousand dollars more cash than the class rule, once, at the freeze. It also strands about forty more jobs a month, every month it runs. Over the twelve months from the freeze on which every alternative is scored, that is roughly 500 stranded jobs against a single release. The release is worth its carrying cost while the stock stays off the shelf: at a quarter of its value a year, about 100 thousand over the horizon. Comparing a one-time release with a per-job price without that step is how a cut that strands five hundred jobs a year comes to look cheap. Crediting the whole 400 thousand inside the year is the most generous reading the cut can get.

Now ask the question. If nobody in the room will put a stranded job below 1,500 dollars or above 6,000, once expedite freight, a held bay, and a conservative goodwill term are counted, the compensation for a year of extra stranded jobs lies between about 700 thousand and 3 million. That whole range sits above the 100 thousand of carrying cost, and above the generous 400 thousand as well. The ranking settles without anyone agreeing on the exact price of a stranded job.

A range settles a ranking only when the amount on the other side lies outside it. Had the release fallen inside the range, the ranking would be genuinely unresolved, and the choices would be to narrow the range with evidence, by pricing a stranded job from the expedite trail and the bay records, or to put the exchange in front of the owner as the one sentence on which the pick turns. Appendix C lays out the Meridian table.

The point is not arithmetic. It is that the judgment the ranking depends on becomes a sentence somebody can contest. Inside a weighting scheme, the objection comes out as “I don’t trust the model.” Written as an exchange rate, it comes out as “that is too cheap for a stranded job,” which is a conversation the room can have.

Dollars, not because dollars are holy

Not every consequence is money. Safety is not a rounding error, a legal constraint is not a cost to trade off in a spreadsheet, and planner time is a real input to whether the commitment survives the floor.

Economics here means consequences in the units in which they are actually felt. Often that unit is money; when it is not, name the unit rather than replacing it with a percentage from the quarterly pack. “Fill up two points” does not say who is spared a stranded job. “About forty fewer vehicle-stopping misses a month than the blanket cut, at this holding cost” does, and only the second can sit beneath a recommended rule as a reason.

The dollars are approximate too. That is not the relevant test; the test is whether the economics rank the named alternatives better than fill and turns do, and even a crude ratio can. If a plausible revision of the ratio would reverse the ranking, record it as a condition on the recommendation, which is the work of Chapter 10, rather than retreating to fill rate.

A test you can run in five minutes

Take the top line of your current write-up. If its unit is a house metric, rewrite it in the unit a person in the room would feel. If two consequences are being traded, check that both are stated over the same horizon, and that the range you would defend for the exchange rate lies entirely on one side of the amount it is traded against.