Inventory optimization is a decision, not a forecast

There is a quiet assumption buried in most inventory optimization projects: that the problem is a calculation. Find the right safety stock. Set the right min-max.

Solve the economic order quantity. Get the math right, and the stock will optimize itself. Software vendors sell to this assumption, consultants build engagements around it. And supply chain teams spend quarters tuning the parameters of models designed for a world that no longer exists.

The math is not wrong. Safety stock formulas, service-level optimization, and reorder-point logic are sound operations research. They have been since the 1950s.

The problem is narrower and more uncomfortable: the calculation is no longer where the leverage is. In modern retail, the formula produces a target — and the target is the easy part. The hard part is the thousands of contextual decisions that surround it, and those are exactly what the formula doesn't make.

This article looks at why inventory optimization, as it is usually practiced, solves a problem retail has largely outgrown. Where the real leverage moved. And what changes when you stop treating optimization as a number to compute and start treating it as a stream of decisions to make.

What "inventory optimization" usually means

Ask three vendors what inventory optimization is and you get variations on one answer: set the right stock level at the right place. Concretely, that decomposes into a handful of well-understood calculations.

Safety stock — how much buffer to hold against demand and lead-time variability, derived from a service-level target and a standard deviation. Reorder point and min-max — the thresholds that trigger replenishment. Economic order quantity — the batch size that balances ordering cost against carrying cost.

Each of these is a formula. Each takes a forecast, a lead time, a cost of capital, and a target service level, and returns a number. The promise of inventory optimization software is that it computes these numbers across tens of thousands of SKU/store combinations faster and more consistently than a human planner with a spreadsheet.

That promise is real, and for stable, high-volume, predictable items it still delivers. The trouble is that the share of a modern retail assortment that fits those assumptions has collapsed. And the formula degrades quietly on everything else.

Why the formula breaks in real retail

The safety stock equation is only as good as its inputs. And in retail, every input it depends on has become unreliable in a way the formula cannot see.

The forecast it depends on is the weakest input

A safety stock number is a function of forecast error. You hold buffer because the forecast is wrong, and the size of the buffer scales with how wrong it tends to be. So the entire calculation rests on a forecast — and on a stable estimate of that forecast's error.

In fast-moving categories, neither holds. Fashion and seasonal items have short histories and high variance — by the time you have enough sales to estimate the error, the season is half over. A forecast carrying 30% error doesn't make the optimization slightly off; it makes the "optimal" safety stock essentially arbitrary, a precise answer built on a guess.

This is the deeper point of why machine learning forecasting isn't enough. Improving the forecast by a few points does not rescue a decision process that treats it as ground truth. The formula has no way to express "I don't trust this number, so let me act more cautiously and re-decide next week." It just computes.

The constraints that decide the move aren't in the model

Even with a perfect target, the formula doesn't tell you what to actually do — because the binding constraints rarely live in the model.

A supplier has a minimum order quantity that overshoots your reorder point. A store's role in the network means it should hold display depth regardless of velocity. A markdown calendar is three weeks out, so reordering now is the wrong call even though min-max says replenish. A transfer from an overstocked store is cheaper than a new purchase order, but the optimization engine only knows how to reorder.

These aren't edge cases. They are the substance of the decision. A min-max system that ignores them produces technically valid numbers that operators override or ignore. It's why so much retail data becomes useless without a decision layer on top of it.

Optimization is a stream of decisions, not a single number

Here is the reframe. The reason classic inventory optimization underdelivers is that it answers the wrong question. It answers "what is the optimal stock level?" when the operational question is "given everything I know right now, what should I do with this SKU in this store today?"

Those are not the same question. The first is a target. The second is a decision — and it has to be made repeatedly, under changing conditions, against constraints the target doesn't capture.

Consider one SKU in one store across a season. Week 3, it's selling ahead of plan — the decision is whether to pull forward replenishment or protect stock for the network. Week 7, it slows in this store but accelerates in two others — the decision is a transfer, not a reorder. Week 11, sell-through stalls — the decision is the markdown timing, weighed against carrying cost.

Not one of those is a safety stock calculation. They are arbitrations — between holding cost and stockout risk, between this store and the network, between marking down now and waiting. A static optimization recomputed each quarter cannot make them. A decision process running at the cadence of the business can — which is exactly the gap between continuous replenishment and the weekly meeting.

What to evaluate differently

If optimization is a decision stream, the questions you ask of a system change.

Stop asking "how accurate is the optimization engine?" and start asking "how many decisions does it actually make. And how many survive contact with operations?" A system that produces mathematically optimal targets that planners override 70% of the time is not optimizing inventory — it is generating ignored numbers.

Ask where the business rules live. If supplier minimums, store roles, and markdown calendars sit outside the model as manual overrides, the system isn't deciding — it's proposing. And a human is doing the real optimization by hand. The rules have to be inside the arbitration, not bolted on after.

Ask whether the system can act. A target that lands in a report still needs someone to turn it into a purchase order, a transfer, or a markdown. The leverage is in closing that gap. The optimization is only as good as the decisions it actually executes. That's the same reason overstock keeps costing more than the ERP shows: the cost accrues in the decisions nobody made in time.

The Solya angle

This is the logic Solya is built on. Not a better safety stock solver bolted onto your existing stack — a decision layer that treats inventory optimization as the continuous arbitration it actually is.

Solya connects to your POS, ERP, and supply chain systems and rebuilds a live SKU/store view of the network. The intelligence layer scans it continuously and frames the real decisions — reorder, transfer, mark down, return. Your business rules are embedded in the arbitration, not applied as a filter afterward. A supplier minimum, a store's display role, a markdown three weeks out: these shape the decision from the inside.

The decisions that clear your rules are then propagated through the orchestration layer into the systems that execute them — no re-keying, no report that waits for someone to act. The same mechanism powers continuous replenishment as a use case rather than a number recomputed each planning cycle.

The point isn't that the math disappears. Targets still matter; safety stock still has a place. The point is that the math becomes an input to a decision, not the deliverable. Optimization stops being a quarterly calculation. It becomes what it should have been all along — a stream of good decisions, made fast, at scale, against the constraints that actually bind.

The question to ask

Look at your last inventory optimization initiative and ask one thing: did it change the numbers, or did it change the decisions?

If it tuned parameters and produced better targets that your planners still override every week, you optimized the formula and left the leverage on the table. The targets were never the problem. The thousands of unmade or late decisions around them were — and those are not a calculation you can solve once. They are a process you have to run, every day, at the size of your network.

---

Is your inventory optimization solving the right problem?

At Solya, we offer retail supply chain and merchandising leaders a 30-minute diagnostic to assess, on your own assortment, whether your inventory optimization is computing targets or actually moving decisions. You'll walk away with:

• A read on where your optimization stops — at the target, or at the executed decision
• An estimate of how many recommendations survive contact with your operations today
• The first SKU/store decision loops worth closing to convert stock into performance faster