When You Think You Need an Inventory Forecast
Several times a month, a team arrives at Lokad with a familiar request: “We need an inventory forecast. Ideally 3 to 12 months out. And, while you’re at it, a forecast of our purchase orders too.”
I always take this request seriously, because it usually signals something important: the company has reached the limits of its spreadsheets and planning rituals, and it is looking for a clearer view of the future. But I also know, from long practice, that this phrasing almost always points in the wrong direction.
In my book Introduction to Supply Chain, I argue that supply chain is, first and foremost, about making better economic decisions under uncertainty, every single day: decisions about what to buy, where to stock it, how to move it, how to price it. A demand forecast, or an inventory projection, is only useful to the extent that it serves those decisions. When the forecast becomes the star of the show, something has gone wrong.
This essay is for the practitioners who feel that urge to ask for “an inventory forecast” or “a PO forecast.” My intent is not to scold you for using the wrong words, but to help you pause and reframe what you truly need—before you invest time, money, and organizational attention in the wrong thing.
The reflex: “Show me the future”
When people say “inventory forecast,” they usually picture a kind of movie of their warehouses and stores for the next 12 months. They expect to see, for every SKU and location, a neatly drawn curve of stock on hand, and alongside it a calendar of purchase orders that will supposedly get them there.
Behind that picture sit a few implicit beliefs:
You assume there exists one “natural” trajectory for your inventory, waiting to be uncovered if only the forecasting technology is clever enough. You assume that if you had this trajectory, aligning teams around it would fix most of your problems. You assume that inventory and PO projections are levers in themselves—objects you can act on directly.
This worldview is not accidental. It is the mainstream supply chain playbook. Traditional approaches begin with plans, targets, and single-number forecasts, and behave as if the world will hold still long enough for those numbers to remain valid. Most planning software, most S&OP rituals, and most textbooks have been built around this picture.
If you have been trained in that world, asking for an “inventory forecast” is not a mistake; it is the natural next step. But it is still the wrong question.
The uncomfortable truth about future inventory
Let us take the phrase “inventory forecast” at face value. What you are asking is:
“Tell me how much stock I will have at each location, in each future period.”
The first problem is obvious, yet often downplayed: the future is irreducibly uncertain. No amount of data or machine learning will buy you a deterministic world. Unpredictable promotions, competitor moves, regulatory shocks, supplier issues, macro events—none of these are going away. Lokad’s own work over the last decade has only reinforced this conclusion: classic time-series forecasts are not just inaccurate; they are incomplete. They squeeze uncertainty into a single number and treat your company as a passive observer, like an astronomer watching planets move.
But there is a second, deeper problem. Future inventory levels do not just depend on the outside world. They depend on decisions that have not yet been made. Future stock positions depend both on irreducible uncertainty and on the purchase orders and allocations you will choose to issue. It is therefore unwise—and quite unprofitable—to assume that a single projection of the future state of inventory can be deemed ‘correct enough’.
There is, in reality, no such thing as the inventory forecast. There is only:
- “If we keep making decisions according to this policy, here is the distribution of possible inventory levels we should expect.”
In other words, inventory in the future is not a natural phenomenon to be forecast. It is the outcome of your policy interacting with an uncertain world. Until you define the policy, you are asking a question with no stable answer.
Forecasting your own decisions?
The situation is even stranger with “PO forecasts.” A purchase order is not something the universe throws at you. It is your own decision: a commitment of cash, capacity, and reputation.
When you ask for a PO forecast, you are essentially asking:
“Please forecast what we will decide to do, several months before we have actually decided it.”
You can, of course, simulate the output of a policy. Once you have a clearly defined decision mechanism—an engine that, when fed data, emits purchase orders—you can simulate many possible futures of demand, lead times, and so on, and see which POs that engine would issue under each scenario. That is a perfectly sensible exercise.
But notice the order of operations: you must have the decision mechanism first. The PO “forecast” is simply one view on how that mechanism behaves under uncertainty. It is not a freestanding object you optimize in isolation.
The mainstream approach often reverses this logic. It starts with a point forecast of demand, pushes it through an MRP or APS system, and treats the planned orders as a kind of “baseline future,” to be tweaked manually through S&OP. When I call this a methodological dead end, I do not mean that nothing ever works; I mean that this paradigm exhausts enormous energy on maintaining a picture of the future, instead of tuning the decisions that actually create it.
Numbers that act, numbers that merely describe
This brings us to a distinction that, in practice, makes or breaks supply chain initiatives: the difference between numbers that do something and numbers that merely describe something.
A purchase order, a transfer order, a price change—these are numbers that act. They commit scarce resources to specific purposes. They move atoms and money. They are decisions in the strict economic sense.
An inventory forecast curve is a description. It does not, by itself, order anything or ship anything. It can help humans make decisions, but only indirectly, by influencing what they choose to do.
When a company frames its goal as “getting an inventory forecast,” it is, often without realizing it, elevating descriptive numbers above acting numbers. The implicit hope is that if the description is precise and widely shared enough, good decisions will somehow emerge.
My experience says the reverse is more reliable: start from the decisions you want to automate and improve, then engineer whatever descriptive numbers are needed to support those decisions. Not the other way around.
At Lokad, this is why we talk about unattended decision engines and why we insist that a serious optimization system must return finalized decisions, not just dashboards that still need rescuing by planners every morning.
Decisions first, forecasts second
What does it mean, concretely, to put decisions first?
It starts with a very simple, but surprisingly rare, question:
“What are the recurring decisions we need to make, and how should we judge whether they are good?”
In most companies, the recurring decisions include creating purchase orders, allocating stock across locations, choosing transport modes, accepting or rejecting customer orders under constraint, and setting or adjusting prices. You make these decisions every day, whether you think of them as “supply chain” or not.
Next, you must decide how to evaluate those decisions in money. This is the economic model: the way you translate stockouts, overstocks, lead-time buffers, transport choices, and so on into a common unit—typically dollars of profit, properly adjusted for risk. Lokad spends a great deal of effort here, because if the pricing of trade-offs is wrong, the optimization will faithfully pursue the wrong goal.
Only then does forecasting come into play. We forecast all the pieces of the world you cannot control but that matter for your decisions: future demand, future lead times, returns, scrap rates, and so forth. We do this probabilistically, not because probabilities are fashionable, but because they are the only honest way to represent irreducible uncertainty.
Those probabilistic forecasts feed an optimization routine that searches, not for a pretty plan, but for the set of decisions that maximizes expected economic return while respecting your constraints. The output is not a demand plan, not an inventory curve, but a concrete list of actions: order 120 units of this item, transfer 40 units of that one, adjust this price upward, and so on.
Once you have this machinery in place, you are finally in a position to ask meaningful “what if” questions about the future.
What projections are good for
At this point, you might object:
“Fine, but my CFO still wants to know where working capital is going. My operations director still wants a view of capacity saturation. Are you saying we should never project inventory?”
Not at all. Lokad absolutely does calculate future inventory levels and purchase orders—once the decision engine exists. Our own documentation says so quite plainly.
The key is that these projections are now anchored in something real: a fully specified decision process plus a probabilistic view of uncertainty. You are no longer pretending that there is a unique, ordained path for inventory. Instead, you are saying:
“If we continue to follow this decision policy, and the world behaves according to these probabilistic patterns, here is the distribution of outcomes we can expect for inventory, capacity, service, and cash.”
You can run multiple scenarios by changing the policy or the economic model. Suppose you increase the implicit penalty for stockouts on a critical product family. The engine will respond by ordering more, earlier. The projected inventory distribution will shift upward; your working capital requirement will rise, but your risk of shortage will fall. You can see this trade-off explicitly, in dollars, before you commit.
In that sense, inventory and PO “forecasts” become instrumentation. They are views on how your chosen policy behaves under uncertainty, not standalone artifacts you fight over in a monthly consensus meeting.
Agency, waiting, and the danger of long rigid plans
There is another reason I am wary of long, rigid inventory and PO forecasts: they quietly erode the agency of the company.
A supply chain is not a one-shot optimization problem. It is a long sequence of decisions made as new information arrives. In my recent essay on sequential decision analytics, I describe two instruments that have proved useful in practice: a “window of responsibility” for judging decisions, and what I call the “economics of waiting.”
The idea of a window of responsibility is simple. You do not need to script an entire twelve-month season to judge whether ordering a container today was wise. You pick a reasonable horizon—say, the next few months—over which you attribute the financial consequences of that order. After that, later decisions take over the responsibility. This keeps accountability sharp without pretending that today’s plan must anticipate every twist and turn.
The economics of waiting is the notion that “do nothing yet” is a fully legitimate option. You act only when the expected, risk-adjusted return of the best move clears your internal cost of capital plus the option value of having more information tomorrow. Waiting preserves options; acting too early freezes them.
Now ask yourself: what happens culturally once a highly detailed 12‑month inventory and PO plan has been blessed by management?
People start treating that plan as a commitment, even if everyone knows, intellectually, that it will be wrong. Requests to change course are perceived as deviations rather than as normal reactions to new information. “Sticking to the plan” becomes an objective in itself, separate from making the best decisions given what you know today.
In that environment, long-range projections do not just inform decisions; they shape them in ways that are very hard to correct. The company slowly trades away its agency—the ability to respond to reality as it unfolds—in exchange for the illusion of a stable future.
A better way to phrase the question
So, if “we need an inventory forecast” is the wrong question, what should you ask instead?
A more fruitful starting point sounds like this:
“We need to improve the quality and automation of our purchasing and allocation decisions, under uncertainty, while keeping full visibility on capital and risk. What would it take to build an engine that does this—and how can we then inspect its behavior over the next 3 to 12 months?”
This phrasing assumes that decisions, not forecasts, are the primary output. It assumes that uncertainty is here to stay, so you will need probabilistic views rather than single numbers. It assumes that financial impact, not proxy KPIs, is how you will judge success.
Once you frame the problem this way, a different project takes shape:
You identify the key recurring decisions. You design an economic model that prices the trade-offs involved. You build or adopt a predictive layer that estimates the relevant uncertainties probabilistically. You implement an optimization layer that turns those forecasts and prices into concrete decisions. You deploy it in such a way that it can run unattended for the mundane cases while surfacing the truly exceptional ones for human review. And only then do you start generating inventory and PO projections as ways to understand and stress-test the resulting policy.
Notice that you may end up with screens that look, from afar, quite similar to what you originally imagined: graphs of stock levels over time, tables of expected purchase orders, scenarios under different assumptions. But these artifacts are now downstream from the real work, not mistaken for the work itself.
A practical pause for practitioners
If you take only one thing from this essay, let it be this:
Decisions are the outcome; forecasts are just tools.
Before you commission a grand “inventory forecasting” initiative, pause and ask yourself a few concrete questions.
Are you clear on which decisions you want to automate or support? Do you know how you will judge those decisions in economic terms, not just in service levels or forecast accuracy? Are you ready to accept that the future is a family of possible scenarios, not a single line? And are you prepared to let a decision engine revise its recommendations as reality unfolds, rather than defending a 12‑month plan against the facts?
If the answer to these questions is no, then a better inventory forecast will not save you. At best, it will give you nicer graphs to look at while you continue to make decisions by habit and negotiation. At worst, it will harden those habits into a rigid script.
If the answer is yes, then you do not “need an inventory forecast” in the traditional sense. You need a supply chain that has learned to think in decisions, in uncertainty, and in money. The projections will follow naturally, and they will finally mean what you hoped they meant all along.
