A vendor of radio, transport and access equipment sells to a small number of operators, and each of them buys through tenders, frame agreements and rollout waves rather than a steady order stream. Shipment history describes programmes that have already ended and says little about the next quarter. The useful forecast is built from the tender pipeline, with the chance of winning and the timing of the volume treated as separate uncertainties, and from call-off behaviour inside contracts already won. Long-lead silicon and optics are committed before most of those decisions are made, so the plan has to separate committing components from committing to build.
The quarterly demand review at an equipment vendor tends to open on the same slide. Three large tenders in the pipeline, two frame agreements in their second year, and a factory asking for a build plan it can staff. Sales wants stock ahead of the biggest award, because the operator will expect first deliveries within weeks of signature. Finance wants the opposite, because the last time the company built ahead of a tender it then lost, the radios sat in a warehouse until a hardware revision made them hard to sell to anyone.
Behind that argument sits a structural fact. An equipment vendor may have a few dozen customers that matter and a handful that account for most of the volume. Each of them buys in programmes: a new spectrum band, a capacity upgrade in the cities, a regional swap of an incumbent vendor, a rural coverage obligation with a deadline attached. Demand arrives as a small number of large decisions made inside other companies, and the planning method has to be built around that rather than around a demand series.
Shipment history describes programmes that have ended
Statistical forecasting assumes the process that generated the history will carry on generating the future. In equipment sales the history is a sum of programmes, each with a start, a ramp, a peak and an end. When a swap programme finishes, volume does not decay along a smooth curve. It stops, and the next programme starts somewhere else, for a different operator, on a different band, often with a different product.
A model fitted to the last eight quarters of shipments will therefore project the tail of a finished programme into a period where none of it applies, and it will miss the next programme entirely because nothing in the history announced it. The accuracy reports look poor for reasons the model cannot fix, and planners learn to ignore the statistical number, which is a reasonable response to a method applied to the wrong object.
History does carry information for part of the book. Spares, expansion cards, power and battery units for an installed base, and the steady build rhythm of smaller operators behave like run-rate demand and respond to normal methods. The first planning step is to split the book in two: programme demand, planned from the pipeline and from contracts, and run-rate demand, forecast from history. Keeping both in one series is how the run-rate forecast gets distorted by a programme that ended and the programme forecast gets smoothed into a ramp no customer ever ordered.
A tender carries two uncertainties that should be kept apart
A tender in the pipeline has a probability of being won and, separately, a shape if it is won: when the award lands, how fast the operator ramps, and what share of the network you get. Multiplying the win probability by the full programme volume produces a weighted number that will never happen. The outcome is either nothing or the programme, and a factory staffed for the weighted middle is wrong in both cases.
Probability weighting works across many small opportunities, where wins and losses average out. It fails for the three or four tenders that dominate the next year. Those are better planned as explicit scenarios, win and lose, with the material and capacity plan for each written down. The decision then becomes what to commit in the base case and what the cheapest hedge is for the win case, which is a question a finance team can price.
Share is its own forecast. Many operators split their radio network between two vendors by region, and a renewal can move a region from one vendor to the other without the operator's total spend changing at all. Forecasting the operator total and applying a historical share hides exactly the event that matters. Forecast by region, with the share in each region treated as a decision that can change at a known contract date.
Timing given a win deserves the same scrutiny. Award dates in the pipeline are usually the date sales expects, and they slip. Measure the gap between the expected award date recorded at each pipeline stage and the actual award date, by customer, and use that distribution when deciding how early to commit anything.
Call-offs inside a won contract follow the operator's build
A frame agreement states volumes, price steps and a term. The orders that actually arrive are call-offs, and they follow the operator's construction programme, which slips for reasons neither party controls: permits, landlords, power connections, civil contractors. The operator's rolling forecast to the vendor is usually its plan, and plans in network rollout carry an optimism bias that varies by operator and by region.
Treat each customer's forecast as an input with a measurable error rather than as demand. Compare what the operator said it would call off twelve weeks ahead with what it called off, per customer and per product family, and apply the observed bias before the number reaches the build plan. Share the measurement with the customer. Operators rarely see their own forecast error from the vendor side, and the conversation tends to improve the next forecast more than any model would.
Commercial terms shape the timing too. Volume price steps can pull call-offs forward as an operator tries to reach the next step. It is common for orders to bunch before an operator's financial year end, when unspent capital budget has to be committed or lost, and to go quiet at the start of the new year while budgets are approved. Those patterns repeat, so they belong in the plan as known shapes, attached to each customer's fiscal calendar.
Variants multiply where the radio meets the spectrum licence
A radio product family splits by frequency band, power class, number of antenna ports, regional approvals and power supply type. Transport and access equipment split by optics, interfaces and enclosure. The operator's spectrum holdings decide the band mix, and those holdings differ by country and sometimes by region inside a country, so a single operator contract can need several variants in proportions that change as sites are designed.
Mix error is where equipment stock ages. The total radio count may be right while the warehouse holds the wrong bands. The planning response is to hold stock at the level where commonality ends. Common platforms, baseband units and enclosures can be built to the programme forecast, while band-specific modules and filters are configured as late as the delivery lead time allows. The buffer sits at the point where one product becomes many, which is where it covers the most possible orders.
Software adds a second lever. Capacity and features enabled by licence key move part of the variety out of hardware, so a single hardware variant can serve several configurations and the decision about which one moves to the day of activation. That shifts work from inventory planning to licence management, which has its own forecasting needs, but it is a better place to hold uncertainty than a shelf of band-specific radios.
Long-lead components are committed before the tender is decided
Chipsets, programmable logic, high-speed optics and power amplifiers can carry lead times measured in quarters, and in tight markets suppliers allocate rather than accept orders. The commitment window for those parts often opens before the tender decision that would justify them. Waiting for the award means missing the first deliveries the operator expects. Committing early means carrying the risk of the tender that is lost.
Two separate decisions are being made, and they deserve separate plans. Component commitment is driven by the pipeline scenarios and by how common each part is across the product range. Build commitment is driven by awarded contracts and call-offs. Long-lead exposure is safest in components shared by several product lines, because a lost tender then leaves parts another product can consume, while parts unique to one tender's configuration should wait for the award wherever the lead time allows.
The real decision calendar is set by cancellation terms. Many long-lead orders become non-cancellable well before delivery, and that is the date the risk is actually taken. Track the exposure by component and by the date each order becomes binding, and review that list against the award dates in the pipeline. It is the most direct view of what a slipped tender will cost.
Service obligations outlive the product line
Equipment contracts carry support commitments that run for years after the last unit ships: repair and return, advance replacement, spares held at operator depots and sometimes vendor-owned stock on consignment. The installed base keeps generating demand for repairs and replacements long after the product has left the price list.
That demand meets component end of life. When a chip used in a radio is discontinued, the vendor faces a last time buy sized against the installed base, its failure rate, the repair yield, and the remaining length of every support obligation that depends on it. Buy too few and the vendor is redesigning a board for a product it no longer sells. Buy too many and the stock is written off when the last contract expires.
Repairs make the spares population a circulating pool. A failed unit comes back, is repaired, re-enters stock and goes out again, so the quantity to hold depends on turnaround time and repair yield as much as on failure rates. Planning it as consumption overstates what needs to be bought.
Where this stops
No model forecasts a tender award. Price, technical scoring, relationships and, in some markets, government decisions about which vendors may supply a network decide the outcome, and none of that is in the data a planning system sees. What planning can do is make the cost of each outcome visible early enough to choose a hedge.
The data limit is the pipeline itself. Stage probabilities are usually set by the sales team, and they carry the same optimism as an operator's rollout plan. Measure stage conversion and award timing by customer and by account owner, and replace the default probabilities with the observed ones before they feed the material plan.
This week, take eight quarters of pipeline snapshots and, for every tender that has closed, compare the volume, share and award date recorded at each stage with what actually happened. The resulting table of conversion and slip by customer is the input the equipment plan has been missing.
More on this in industry on the DemandLabs platform.
