Complexity Must Pay Rent

The weak version of an analysis starts from the method the team is known for. To a machine-learning specialist the problem looks like a prediction, to an optimizer it looks like a mixed-integer program, and to an experimenter it looks like a test. The strong version starts from the claim the ranking depends on and asks for the cheapest method that can settle it.

The choice is indifferent to anyone’s preferred method, which is why complexity is a cost that has to earn its place. A method pays its rent only if it changes the ranking of the alternatives, the risk attached to that ranking, or the cost of running the rule. And the rent has to be paid in the conditions where the rule will run, which are rarely the conditions of the notebook.

The ladder, and why you start at the bottom rung

Work from the bottom. Climb only when a lower rung cannot settle a claim that is actually load-bearing.

A conversation on the floor. What does the planner already pad? What does the dealer do when the system says no? The incumbent rule includes its undocumented parts, and if the unofficial rule already differentiates items, you have an existence proof and a challenger in one afternoon.

A count, a slice, a query. Is the damage concentrated? This is where a ranking most often dies or lives. A Pareto is not unsophisticated; it is frequently the claim itself.

A back-of-envelope economic model. What does a miss cost, what does a unit of stock cost, and are those numbers in the same universe? If they are not, a differentiated rule is already plausible before anyone fits a demand model.

A simple statistical relationship. Is there something stable enough to act on, and does it change an alternative?

An experiment, when you can actually intervene and what you would learn is decision-relevant. A forecast, as an input to a rule and never as the product; a forecast that does not change an order is a diagnostic with a press release. A simulation, when the system’s dynamics matter and averages lie. An optimization model, when the alternatives are allocations of scarce things under real constraints a heuristic would silently violate. Machine learning, when a prediction is a load-bearing input to a named rule. Generative models, when the bottleneck is genuinely language or search through messy artifacts.

Each rung upward buys something and charges rent for it. The question at every rung is not whether the method is defensible. It is which claim the rung below could not settle. More than one rung can be used, but a higher rung should never be used to avoid the embarrassment of a lower one.

Exhibit 1Start at the bottom rung and climb only when the one below cannot settle the claim

Why specialists climb anyway

People are hired as the optimization specialist or the machine-learning specialist, and they are asked for what they are. Prestige attaches to the upper rungs while decisions tend to be settled on the lower ones. The specialist who can say when not to climb becomes more valuable for it, not less.

The same habit shows up on the page. “We used gradient boosting” and “missed-job cost is concentrated” do not belong in the same list, because the first describes a method and the second states a reason. Methods go in the appendix; a middle layer made of methods describes the team’s week rather than the ranking.

The rent check

The test for climbing is one sentence. If this method works, which alternative moves, by how much, and in what units? If you cannot write the sentence, you are buying complexity with no lease.

Enough means the ranking is stable under the uncertainty that remains, or you have named the uncertainty that would flip it and decided to proceed anyway. At Meridian Motors, a Pareto of missed-job cost against inventory dollars will probably do more for the ranking than a new forecasting stack, which would be justified only if the ranking turns out to be sensitive to forecast shape in a way a simpler method cannot capture.

Heuristics are first-class

A rule a planner will run beats an optimal solution the planner will quietly override. It is mathematics with an adoption constraint, and if an elegant model survives only by being ignored, the heuristic is the real rule. A method the planner cannot restate will be replaced by a pad, the pad becomes the rule, and the incumbent is back.

The productive sequence is often staged. At Meridian Motors, a three-class rule a planner can explain to a dealer comes first. A full optimization of every item every day might later earn its place, but only if someone can show that class assignment is what the ranking gets wrong. “We could do better with a MIP” does not show that. “Class boundaries misrank a set of items whose missed-job cost would change the pick” would, because it can be tested.

The burden of proof depends on the coupling

Where decisions are loosely coupled, each item or location nearly independent of the others, and the inputs are noisy, the burden of proof lies on the more elaborate method, which has to show that its advantage survives the assumptions it makes about the operation. Where decisions are tightly coupled through constraints that reach across time and place, such as shared dock capacity or a frozen production sequence, the burden shifts to the heuristic. A rule that makes locally sensible choices one item at a time cannot see that structure, and it has to show that its plans are feasible before its simplicity counts for anything. Complexity beyond what the claim needs is paid for with someone else’s freeze date, and so is simplicity that has never been tested against the coupling it ignores.

Get the answer a second way

Whatever rung you land on, reach the answer once more by a different route, and cheaply. Before the optimizer runs, work out roughly what a sensible three-class split ought to release: take the slow movers, estimate what fraction of the holding cost sits in items that never stop a vehicle, apply a rough reduction to their days of supply, and multiply. If the envelope says roughly six hundred thousand dollars and the model says four million, one of them is badly wrong, and you found out in twenty minutes rather than in the meeting.

The check catches the serious defects, a unit confusion, a join that duplicated rows, a constraint that never binds, because they tend to be wrong by an order of magnitude. And if you cannot approximate your own model by hand, you do not yet understand what drives it. If the crude calculation and the elaborate one agree closely and rank the alternatives the same way, the elaborate one has not earned its keep on this decision. Say so.

“Simple will not get past review”

Sometimes that is true. It is a political fact rather than a methodological one. The network can go in the appendix while the top sentence remains the decision the Pareto ranked. If the network implies a different ranking, the problem is not the review but the work, which now holds two incompatible answers.

The test

Take the most sophisticated method in your current project and write its rent sentence: which alternative moves, by how much, in what units. Then write the envelope estimate of the same number. If you cannot write the first, the method goes to the appendix. If the second agrees with the model and ranks the alternatives the same way, the method goes to the appendix too, and the recommendation gets simpler.