Inventory forecasting for distributors

Industry: wholesale distribution. Roles: inventory planner, buyer, operations leader. Capabilities: predict; optimization is a separate decision layer. Human control: planner-approved replenishment. Benefit: potential working-capital improvement and better service. Costs: data preparation, modeling, integration, and ongoing review.

The decision

An operations leader wants to know whether a better demand forecast would let a distributor hold less inventory in a defined product segment without losing availability. The first useful boundary is a forecast that feeds a proposed replenishment quantity, which a planner reviews and approves. Supplier negotiation, allocation between customers, pricing, and the decision to stock a new item remain outside this workflow.

The unit of work is a replenishment decision for one item at one location on one review date. The unit of success is not a low forecast error. It is an approved order that, under the agreed policy, keeps the item available at the target level with less inventory or less expediting than before. A method with a better error score can still produce worse orders if the policy that converts forecasts into purchases ignores lead times, pack sizes, or uncertainty.

Unlike the drafting and extraction chapters, the output here is a number about the future, and errors surface weeks later as stockouts, overstock, or emergency freight. Read this chapter to decide whether a forecasting trial is worth running, how to test it honestly, and what would justify keeping the existing method.

All company names, items, demand histories, forecasts, inventory positions, budgets, and pilot thresholds below are invented teaching examples. They are not customer records, measured performance, vendor results, or industry benchmarks. The candidate forecasts in the worked example are authored numbers standing in for a model's output; no model produced them. The research section describes external sources separately.

Map the work before choosing a method

In the illustrative baseline, a planner reviews items on a fixed schedule. For each item, the planner looks at recent sales, on-hand stock, open purchase orders, supplier lead time, minimum order quantities, and any known upcoming events. An ERP reorder rule may propose a quantity, and the planner accepts, edits, or overrides it. The buyer places the order and handles supplier exceptions.

Record which forecast and which rule were actually available on each decision date. Without that record, a later improvement cannot be attributed: a better forecast, a changed safety buffer, and a planner who became more cautious all reduce stockouts.

Workflow map

Demand history → exception flags → forecast with range → replenishment policy → proposed order → planner review → buyer places order → outcome recorded

At exception flagging, a period with a known stockout or a one-off bulk order is marked rather than silently used as ordinary demand. At forecasting, an item with too little history, or a new item, leaves the automated path for planner judgment. At review, an item whose forecast moved sharply, whose supplier lead time changed, or whose proposed order exceeds an agreed value goes to the exception list. If the forecasting service fails, the existing ERP rule remains available. No branch grants the system authority to place a purchase order.

The map asks two separate questions: is the forecast good, and is the order good? An accurate forecast still yields a wrong order if the pack size or lead time is stale.

Compare against simpler methods

Before evaluating any sophisticated model, fix the comparison set. Forecasting: Principles and Practice describes the mean, naive (last observed value), seasonal naive, and drift methods, and notes that one of these will sometimes be the best method available, though in many cases they serve as benchmarks. FPP3, section 5.2 Add the method the distributor already uses, such as a moving average in the ERP.

As a proposed practice, compare three configurations on the same items and dates: the current method and policy; a simple benchmark with an improved, explicit policy; and the candidate method with the same improved policy. This separates the value of better forecasting from the value of fixing the ordering rule, which may be the cheaper improvement.

Walk through one item family

For your operations lead

The fictional distributor in this exercise is Larkfen Facility Supply, which sells janitorial and safety products to schools and facilities managers. The item is LF-GLV-M, a case of medium nitrile gloves. Demand is measured in cases per week. The supplier lead time is two weeks, the planner reviews the item weekly, and the supplier ships in multiples of 10 cases. All figures are invented.

Input: demand history

Assume every week had stock available, so recorded sales equal demand. That assumption fails in any week the item was out of stock; the data section explains why.

Evaluate three methods on a rolling origin

Weeks 1–8 are the training period. For each of weeks 9–12, produce a one-week-ahead forecast using only data up to the previous week, then compare it with what happened. This is time-series cross-validation, sometimes called evaluation on a rolling forecasting origin: each training set contains only observations before the test observation, so no future data enters the forecast. FPP3, section 5.10

Errors are actual minus forecast. The naive forecast is the prior week's demand. The moving average for week 9 is (46 + 40 + 44 + 48) ÷ 4 = 44.5, and so on. The candidate column is authored.

The scaled error divides each method's mean absolute error by the average absolute one-step naive error on the training data. Here that denominator is the mean of the seven week-to-week changes in weeks 1–8: (4 + 6 + 4 + 4 + 6 + 4 + 4) ÷ 7 ≈ 4.571 cases. FPP3 explains that a scaled error below one comes from a better forecast than the average one-step naive forecast on the training data, and above one from a worse forecast. FPP3, section 5.8

Four weeks of one item prove nothing; the candidate looks best because its numbers were authored. The table shows the mechanics, not a result.

Convert the forecast into a proposed order

The proposed policy for this exercise is an order-up-to rule, chosen for arithmetic clarity rather than as a recommended inventory model:

• Order-up-to level = weekly forecast × (lead time + review period) + safety buffer.
• Inventory position = on hand + on order − backorders − stock allocated to open customer orders.
• Proposed order = order-up-to level − inventory position, rounded up to the supplier multiple.

At the end of week 12, Larkfen has 60 cases on hand, 50 on order, and no backorders or stock allocated to open orders, so the inventory position is 110. Where open customer orders exist, on-hand stock overstates what is free to cover forecast demand, and leaving them out understates the proposed order. The planner has set a safety buffer of 12 cases. That buffer is an invented planning choice, not a computed service-level result.

A 1.5-case weekly difference becomes a one-pack difference after rounding. Whether the extra pack is wise depends on the cost of a glove stockout for a school customer against the cost of holding ten more cases: a business decision the forecast cannot make.

Expected planner package

A useful proposal shows the planner the point forecast (47 cases per week) with its uncertainty range, the policy inputs and the date each was last confirmed, the inventory position with expected receipts, any flags on the history, and the proposed order awaiting approval.

FPP3 notes that point forecasts can be of almost no value without accompanying prediction intervals, and that intervals typically widen as the horizon lengthens. FPP3, section 5.5 Showing the range lets the planner see when a proposal rests on a highly uncertain forecast.

Reject a plausible but wrong use of the forecast

Forecast for next week is 47 cases, so order 47 cases.

This treats the forecast as the order, ignoring the 110 cases on hand or on order, the lead time, the pack multiple, and the buffer. The error is not in the forecast but in skipping the policy. A screen that presents only the forecast invites it.

A second rejected use: the planner reads "47" as a commitment and tells sales that 47 cases will be available for a customer next week. A forecast estimates demand; availability depends on stock, orders in transit, and competing customers.

Change one input at a time

One-off bulk order. In week 16, a customer opening a new building orders 120 cases once, on top of ordinary demand. If that week's total enters the moving average as ordinary demand, it raises each of the next four moving-average forecasts by 120 ÷ 4 = 30 cases per week, inflating orders for a month. The expected behavior is to flag the order as one-off, keep it out of the baseline series, and plan it separately; no method can know it will not recur unless someone records that.

Intermittent item. LF-GLV-XS, extra-small gloves, sold 0, 0, 3, 0, 0, 0, 2, 0, 0, 4, 0, 0 cases in weeks 1–12: nine zero weeks out of twelve. Its four-week moving average at week 13 is (0 + 4 + 0 + 0) ÷ 4 = 1 case per week, a quantity it never actually sells. Percentage errors are infinite or undefined when actual demand is zero and extreme near zero, so they cannot score this item. FPP3, section 5.8 FPP3 describes Croston's method, which forecasts non-zero demand sizes and the intervals between them separately, and cautions that it does not properly handle count data, is known to be biased, and has no algebraic prediction intervals. FPP3, section 13.2 For items like this, the proposed practice is to judge the policy by availability and inventory held, not by point-forecast accuracy.

Supplier lead time doubles. The supplier moves LF-GLV-M from two to four weeks. With the same candidate forecast and buffer, the order-up-to level becomes 47 × 5 + 12 = 247, the requirement becomes 247 − 110 = 137, and the proposed order rounds to 140. The lead-time change moves the order by 90 cases; the choice between the two forecasting methods moved it by 10. The buffer, held fixed here, would probably need to grow as well. When lead times are unstable, fixing lead-time data may matter more than improving the forecast.

Design the data and system

For your technical lead

Forecasts can only be as good as the demand history. As a proposed minimum, assemble stable item identities (with supersession links when a product is replaced), dated orders and shipments, stock availability by period, returns, promotions and price changes known at the time, customer one-off flags, and supplier lead times as quoted and as received.

Sales are not demand. In a hypothetical period with ten units requested and six in stock, recorded sales of six omit four unfilled units if there are no backorders. A method trained on those sales learns that the item sells less than it does, can recommend smaller orders, and can cause more stockouts. Record stockout periods and unmet requests where observable, and flag those periods rather than inferring missing demand from sales alone.

Segment the items. Stable high-volume items, intermittent items, seasonal items, and new items behave differently and may need different methods and different evaluation. Evaluate at the level used for replenishment, for example item by location. A good aggregate forecast can hide damaging item-level errors.

Separate forecast from policy. Keep the service target, safety buffer, pack sizes, minimums, and lead times in an explicit, owned policy layer. When an order is wrong, the team can then see whether the forecast, a policy parameter, or stale master data caused it.

Order by supplier, not only by item. Purchases go out as supplier purchase orders, and supplier minimum order values, prepaid-freight thresholds, and quantity-break pricing can pull an item's order forward or push it back. An item-level policy simulation that ignores these joint constraints misstates order counts, timing, and average inventory. As a proposed practice, simulate at the supplier level where those constraints bind, or state in the results that the simulation is item-level and how supplier constraints were handled.

Define service level. "Maintain availability" needs a number and a definition agreed before the trial, such as the share of demand filled from stock or the share of review cycles without a stockout; the two lead to different buffers.

Version everything. For each decision date, store the history snapshot, flags, method version, forecast, range, policy inputs, and proposed and approved orders; otherwise a backtest will quietly use later corrections. Customer-level order histories can reveal commercial relationships, so decide who may see them and how long snapshots are kept.

Choose methods against this task

For your analyst

Candidate approaches fall into three broad families. Classical statistical methods, such as exponential smoothing and ARIMA, can be fitted to each series separately. Machine-learning methods, such as gradient-boosted trees, can be trained across many items at once and given features like promotions and calendar effects. Pretrained time-series foundation models are designed to forecast new series with little or no task-specific training. These are orientation, not performance claims: this chapter ranks none of them, and none has been benchmarked on distributor data for this guide.

Test a complete configuration: method and version, features, training window, segmentation, and policy. Compare configurations on the business outcome under the same policy, and include operating effort: a model that needs weekly retraining by a data scientist costs more than an ERP setting.

Use the model-selection guide for the structure of a dated decision record; its evaluation method applies, though its examples concern language models. Language-model specifications say nothing about forecasting accuracy. A language model may help a planner read a forecast report, but a fluent explanation is not evidence that the numbers predict demand well.

Work through the economics

For your finance lead

Consider one invented inventory example. Suppose a pilot proposes reducing average inventory in the segment from $500,000 to $450,000 while maintaining the agreed service level. The $50,000 difference is potential working-capital release: a one-time reduction in cash tied up in stock, realized only as purchases fall below what they otherwise would have been. It is not a recurring annual saving.

The recurring effect is lower carrying cost on the smaller balance. At an assumed 20% annual carrying rate, $50,000 × 20% = $10,000 of annual carrying value. A carrying-rate assumption can bundle different things, such as storage, handling, insurance, shrinkage and obsolescence, and the cost of capital; ask finance what the local rate includes. Some of these are cash expenses; others, particularly the cost of capital, are opportunity costs. If annual modeling and operating cost is $6,000, the illustrative net annual carrying value is $4,000 before setup.

At a 20% rate, the recurring operating cost requires a sustained reduction of $6,000 ÷ 20% = $30,000 in average inventory just to break even on carrying value. That excludes setup and any service-level change.

Keep service risk on the same page

Inventory can always be reduced by accepting more stockouts. A result that lowers inventory while lowering availability is a policy change, not a forecasting gain. Before valuing any reduction, confirm that the agreed service measure held, and record lost sales, emergency freight, customer credits, and write-offs separately. This chapter assigns them no value. Do not count one missed order twice, as both avoided expedite spending and avoided lost margin.

Planner time may rise or fall with the exception load; measure it on the baseline's basis and keep it separate from inventory value.

Finance should determine how carrying costs, write-downs, and system costs are classified under the applicable reporting framework and company policy. The economics chapter explains why working-capital release, recurring carrying value, and accounting classification are separate questions. This example is a decision model, not an accounting determination.

Define failure severity before the trial

For whoever runs the pilot

The categories below are proposed pilot rules. Owners should adapt them and agree on them before seeing results.

Track errors caught at planner review separately from those that reached a supplier order; excellent planners can otherwise conceal a poor method. Follow each decision until demand is realized, because forecasting failures surface weeks later.

Build the evaluation

For whoever runs the pilot

Backtest on a rolling origin. Reconstruct the data as it existed on each decision date and forecast forward over the horizons that matter for ordering, at least lead time plus review period. The time-series cross-validation procedure can be extended to multi-step forecasts. FPP3, section 5.10 Include seasonal peaks and disrupted periods, and keep a final interval untouched until the method is chosen.

Prevent leakage. Possible leaks include stockout flags recorded retroactively, promotions entered into the system after they occurred, master data corrected later (item merges, pack sizes), and lead times replaced by averages computed over the full history. Each makes a backtest look better than live use will be.

Choose metrics deliberately. FPP3 explains that errors measured in units cannot be compared across series with different units, which favors scaled measures when comparing items; that minimizing mean absolute error leads to forecasts of the median, while minimizing root mean squared error leads to forecasts of the mean; and that percentage errors fail when actual values are zero or near zero. FPP3, section 5.8 As a proposed practice, report for each segment:

• Mean absolute scaled error, by horizon.
• Bias (mean error), because a small but persistent bias accumulates into stockouts or overstock.
• Prediction-interval coverage: how often actual demand fell inside the stated range.
• Simulated policy outcomes: availability under the agreed definition, average inventory, and number of orders, using the same policy for every method.

Simulate the policy, not only the forecast. Replay each method through the same replenishment policy with historical lead times and constraints. Simulation is not proof of realized savings: real suppliers are late, planners override, and demand responds to availability.

Use several denominators. Report item-weeks evaluated, items per segment, stockout events per segment, and exceptions routed. Keep intermittent and new items separate from stable items instead of averaging them into one score.

For an initial offline gate, a team might require the candidate to beat the best simple benchmark on scaled error for stable items, show no worse bias, and meet the service target in policy simulation with lower average inventory, while every critical item passes simulation without a stockout. These are illustrative management thresholds. A backtest with no observed stockouts does not establish that none will occur.

Run a supervised pilot and decide

Begin in shadow mode: the candidate method produces forecasts and proposed orders alongside the existing process, but planners order as before. Compare the two sets of proposals against realized demand. Then move to a bounded segment, for example stable items from two suppliers, where planners approve candidate-based proposals and the prior method remains available.

Log every override. Record the original proposal, the planner's change, and the reason. Comparing the accuracy of the planner-adjusted forecast with the unadjusted one shows whether adjustments add value. Overrides for known events, like a school closing for renovation, are information the model lacked; habitual rounding up may be protective bias that shows up as inventory.

Compare fairly. Where practical, assign comparable items to each method rather than letting planners choose; otherwise record the selection and treat results as directional. Schedule the decision only after enough replenishment cycles have closed.

The operations leader owns expansion. The planner owns policy parameters and overrides. The analyst or data owner owns the pipeline, snapshots, and backtests. The buyer owns lead-time data. Finance reviews the carrying-rate assumption and any claimed cash effect. Before expanding, complete this checklist:

• Service measure, target, segments, and severity rules were set before results were seen.
• Decision-date snapshots exist and backtests use only data available at each date.
• Stockout, promotion, and one-off periods are flagged in the history.
• The candidate beat simple benchmarks and the existing method under the same policy.
• Intermittent and new items were evaluated separately or excluded.
• Bias and prediction-interval coverage were reported alongside error.
• Override reasons were recorded and analyzed.
• Any inventory reduction was measured with service held at target.
• Working-capital release and annual carrying value were reported separately.
• A reversion path to the prior method exists and has an owner.

Expand only to tested segments. Revert affected items if availability for critical items falls below the agreed threshold.

People and skills

Planners shift from producing forecasts to reviewing exceptions and supplying what the model cannot observe: one-off orders, known events, supplier changes. They need to read a forecast range, recognize bias, and treat forecast and order as separate decisions, with freedom to override and an obligation to explain why. Buyers become the source of truth for lead times. Analysts maintain the pipeline, flags, and snapshots. Leaders should judge the process across many decisions, not the most recent stockout.

Who you need

This pilot needs the roles below, named by responsibility rather than job title. Some can come from outside for the build, but whatever the business owns after go-live, including process ownership, the policy layer, evaluation, and order approval, stays in-house.

Staffing an AI pilot sets out the method behind this table, and Choosing and working with an AI partner covers bringing in outside help.

When to defer

Defer sophisticated forecasting when item master data is unreliable, stockouts are frequent but unrecorded, lead times vary more than demand, or the segment is small enough that planner judgment is cheap. For intermittent items, a simple policy with an explicit buffer may be the better choice.

The simpler alternative may be to fix the policy layer first: confirm lead times, set explicit buffers by segment, correct pack sizes, flag one-off orders, and compare the ERP rule with a seasonal naive benchmark. If that closes most of the gap, the case for a new model may disappear.

What external research establishes

For your analyst

Forecasting: Principles and Practice, third edition, by Rob J Hyndman and George Athanasopoulos (online edition, accessed September 30, 2026), supplies the methods and evaluation concepts cited in this chapter: simple benchmark methods, prediction intervals, accuracy measures including scaled errors and the limits of percentage errors, time-series cross-validation, and Croston's method for intermittent counts. These are methodological references; they report no inventory outcomes for any distributor.

Forecasting competitions have compared statistical and machine-learning methods on retail datasets, but this chapter does not rely on them. Competition results were not reviewed; any finding would need to be read in full and scoped to its setting.

Evidence still needed and related reading

Obtain an attributable distributor account of the current planning process, decision-date snapshots, stockout records, lead-time variability, and override practice. Measure service levels, inventory, expedite spending, and planner effort before and during a trial. No operator has been contacted, no historical data has been analyzed, and no model has been benchmarked for this chapter.

Continue with wholesale distribution, operations leaders, finance leaders, model selection, pilot design, and the economics worksheet. For the neighboring distributor workflows, see order intake and catalog enrichment.

Companion templates

Prepare your pilot with the pilot kit, evaluation scorecard, and evidence checklist. The companion workbook provides blank economics inputs and case records; worked chapter figures remain illustrative examples.