AI Demand Forecasting for Nigerian Businesses: Units, Lead Times and Lost Sales

Manufacturers, importers, distributors and multi-branch retailers all ask the same question in different words: how much should we make, bring in or move, and where? A sales forecast cannot answer it, because sales only record what you managed to supply. If a product was out of stock for nine days last December, your sales history says demand collapsed. It did not; you did.

Demand forecasting exists to separate those two things. It is the forecast that drives production plans, import orders, transfers between branches and safety-stock levels — the decisions where being wrong is expensive in both directions.

What is demand forecasting, and how is it different from sales forecasting?

Demand forecasting estimates the quantity customers want; sales forecasting estimates the quantity you will sell and invoice. In a market where stock-outs, delivery failures and supplier delays are routine, the gap between the two is often the most valuable number in the business.

A distributor with a 10-week import lead time must forecast 10 weeks ahead plus a review cycle. A baker with a one-day ingredient lead time forecasts tomorrow. The lead time, not the accounting calendar, sets the horizon.

The lost-sales problem and how to fix it

When a product is unavailable, recorded sales fall to zero while real demand continues. Statisticians call this censored demand. Train a model on censored history and it learns to under-forecast exactly the products that sell out — the fastest movers, where under-forecasting costs most.

Practical fixes, in order of ease:

• Record availability. Add a daily or weekly flag per SKU per location: in stock, out of stock, partially available. This single field improves forecasts more than any modelling technique.
• Log unfulfilled orders. When a customer asks for something you do not have, capture it — in the POS, in the order sheet, or as a one-tap entry in the sales app. Many Nigerian businesses have this information in WhatsApp conversations and nowhere else.
• Estimate during known gaps. For historical periods you cannot recover, treat stock-out days as missing data rather than zero demand, and let the model interpolate from comparable periods and similar products.
• Watch substitution. When customers buy the 75cl instead of the 50cl because the 50cl is finished, the smaller product's demand is understated and the larger one's is overstated. Record substitutions where you can, or model product families together.

Choosing the grain and the horizon

Forecast at the grain at which you make decisions, and no finer.

• SKU by location by week suits multi-branch retail and distribution with weekly replenishment.
• SKU by month suits importers with long lead times and monthly ordering.
• Product family by week, split by historical share works when individual SKUs are too sparse to model — a common situation in Nigeria, where catalogues are long and tails are thin.

A useful discipline is hierarchical forecasting: forecast at several levels (total, category, SKU) and reconcile so the numbers add up. Totals are far more stable than individual items, and reconciling pulls noisy SKU forecasts toward a sensible aggregate.

Horizon: forecast at least as far as the longest lead time in the chain, plus the ordering cycle, plus a safety margin. For an importer using sea freight, that is frequently three to four months; for that horizon, present a range wide enough to be honest.

Which forecasting method fits your business?

Most Nigerian businesses should start at the top of this table and move down only when a measured accuracy gap justifies it. A gradient-boosted model over 600 SKUs is worth building when your top 100 SKUs are already well forecast and the tail is costing money.

Forecasting new products and promotions

Two situations defeat naive models, and both are common.

New products have no history. The workable approach is analogue forecasting: identify similar existing products by category, price band and pack size, use their launch curves as a starting shape, and adjust for planned distribution. Review weekly for the first six to eight weeks and replace the analogue with real data as soon as the pattern establishes. AI helps by finding the analogues objectively rather than by a manager's memory.

Promotions change everything temporarily. A price cut, a bundle, a Jumia or Konga campaign day, or a distributor incentive produces an uplift, then usually a dip as customers who stocked up stay away. Model the uplift explicitly as a feature with the promotion type, depth of discount and duration, and model the post-promotion dip too. Forecasting promoted weeks as if they were normal is one of the largest single sources of error in Nigerian FMCG planning.

What data you need before you start

• Two years of transaction-level sales in units, by SKU and location (12 months is the absolute minimum).
• Availability or stock-out records per SKU per location, however coarse.
• Price history per SKU, with every change dated.
• Promotion and campaign records: what, when, how deep, which channel.
• Product master data: category, pack size, brand, whether it replaced another item.
• Supplier and production lead times, with their variability — not just the quoted figure.
• A Nigerian event calendar: public holidays, Ramadan and both Eid festivals, Easter, December, school resumption, salary week, market days where relevant.
• Known disruptions: port delays, road closures, fuel scarcity, strikes, bank downtime, flooding.
• Distribution changes: new branches, new distributors, lost accounts, listings won or lost.

What changes for Nigerian businesses

Lead times are long and variable, so forecast the lead time, not the month. An importer who forecasts January demand in December has already missed the decision. Clearance delays at the ports, documentation and haulage mean the variability of the lead time matters as much as its average. Record actual receipt dates, not promised ones.

Price changes are frequent, so model units with price as a driver. When a pack price rises 25% in a year, a units model with price elasticity learns the demand response. A naira model simply shows growth that never reached the warehouse.

Sell-in is not sell-out. Manufacturers and importers see orders from distributors, which reflect distributor stocking behaviour and credit position as much as consumer demand. Where possible, collect sell-out data from key distributors or retail partners; forecasting on sell-in alone amplifies the bullwhip effect and leaves you overstocked after a heavy trade-loading month.

Informal channels carry real volume. Open markets, kiosks, roadside vendors and WhatsApp resellers may be a large share of demand and are poorly recorded. Sampling a set of these customers with a simple order form or field-sales app gives the model a signal it otherwise lacks.

Seasonality is layered and religious calendars shift. December and the festive period, Ramadan (which moves roughly eleven days earlier each year and changes daytime consumption sharply), Eid-el-Fitr and Eid-el-Kabir, Easter, back-to-school in September and January, rainy and dry seasons, and harvest timing. Supply these as calendar features; do not expect the model to find them.

Power and connectivity shape the record. Where POS terminals or the inventory system are offline for hours, sales get written in a book and entered later, or not at all. A demand forecast is only as good as the record, so offline-capable capture is a real design requirement, not a nicety.

Example (hypothetical): a household-goods importer supplying the South-West

This is an illustrative scenario, not a Linestech client.

An importer of kitchenware and plastic housewares brings containers in roughly every ten weeks and supplies about 90 wholesale customers across Lagos, Ibadan and Abeokuta, plus a small online store. It carries around 450 SKUs. Orders are placed by the managing director based on what the warehouse says is finishing.

The result is predictable: fast movers run out six weeks after each container lands, while slow lines sit for two years and tie up working capital that the business borrows at cost.

The build:

1. Twenty-four months of invoices are exported in units by SKU and customer, and mapped to categories and pack sizes.
2. The warehouse team reconstructs availability history from bin cards and the stock ledger; from that point a weekly availability flag is captured in the inventory system.
3. Price changes, container arrival dates and past trade promotions are compiled into a driver table.
4. Baselines are computed: last-year-plus-growth per SKU, and a three-month moving average.
5. SKUs are split — the top 120 by volume get individual time-series models; the tail is modelled as families with share-based allocation; genuinely intermittent items use an intermittent-demand method.
6. Forecasts are produced at a 14-week horizon (10-week lead time plus review and buffer), as ranges, reconciled up to category and total.
7. The output is a container planning sheet: recommended quantity per SKU, with a confidence band and a flag for items where the forecast is unreliable and judgement is needed.

Practical effect: the order is built from a defensible number rather than from the last complaint, slow lines are cut deliberately, and the business can see which SKUs it consistently under-orders because they sell out early.

How much does AI demand forecasting cost in Nigeria?

Indicative 2026 ranges; actual quotes vary with catalogue size, data condition, vendor and exchange rate. Always compare two or three written quotations on identical scope.

Separate recurring costs: cloud hosting for scheduled jobs, any planning or BI subscriptions billed per user per month in US dollars, and data storage. Because the USD-denominated items move with the exchange rate, review them quarterly rather than annually.

Step-by-step: building a demand forecast that people use

1. Fix the decision first. "The container order placed every ten weeks" or "the Tuesday branch transfer list."
2. Set the horizon from the lead time, including its variability, not from the financial calendar.
3. Start recording availability today, even on paper, while you assemble history.
4. Export and clean two years of unit sales by SKU and location; map products properly, including replacements and renames.
5. Compute naive baselines and their error. This is the bar every model must clear.
6. Segment the catalogue into fast, slow and intermittent, and assign methods accordingly.
7. Add drivers: price, promotions, the Nigerian event calendar, distribution changes.
8. Backtest on held-back months and report error by segment, not just overall.
9. Produce ranges and agree the ordering rule: order to the upper band for fast, high-margin items; to the lower band for slow or perishable ones.
10. Run in parallel with the current method for one full ordering cycle.
11. Automate the feed and the schedule, and deliver the output inside the planning tool the buyer already uses.
12. Review every cycle: accuracy by segment, SKUs to move between methods, drivers worth adding.

How to judge the forecast, and what accuracy to expect

Measure error on held-back data and then continuously, always against a naive baseline, and always by segment.

• Fast-moving, established SKUs forecast well at weekly and monthly horizons and usually beat baselines clearly.
• Slow and intermittent items forecast poorly by nature. For these, the right output is a stocking policy — a reorder point and safety stock — rather than a point forecast.
• New products should be reported with explicit uncertainty for their first two months.
• Promoted periods should be scored separately; if promotion weeks dominate your error, the promotion features need work.

Two supporting measures matter as much as error size: bias (are you consistently over or under?) and service level achieved (did the forecast actually prevent stock-outs?). A forecast with moderate error and no bias is more useful operationally than a lower-error forecast that is always short.

Mistakes to avoid

• Treating stock-out zeros as demand. The single most damaging mistake; it teaches the model to starve your best products.
• Forecasting naira instead of units. Price inflation masquerades as growth, and the warehouse cannot receive naira.
• Ignoring the lead time. A four-week forecast is useless to a ten-week supply chain.
• Modelling every SKU individually. Long tails need family models and stocking policies, not 450 bespoke forecasts.
• Leaving promotions out of the model. Promoted weeks then contaminate the baseline pattern.
• Using distributor orders as consumer demand. Trade loading and credit cycles distort the signal.
• No human override path. Planners know about the contract you just won; give them a documented way to adjust, with the adjustment recorded and scored.
• Buying software before fixing records. No tool compensates for missing availability data.

Conclusion

Demand forecasting pays for itself when it is built on units rather than naira, at a grain that matches the ordering decision, over a horizon set by the real lead time, with stock-outs recorded so the model learns actual demand rather than your supply failures. Start with baselines, segment the catalogue, model promotions and new products deliberately, produce ranges, and track bias and service level alongside error. Indicatively, a pilot starts from around ₦500,000 and an automated planning system from around ₦2,000,000 in Nigeria, plus recurring hosting and subscriptions. The biggest single improvement most Nigerian businesses can make costs nothing: start recording what was out of stock.

If your purchasing, import or production planning runs on memory and last-minute pressure, Linestech can help you assess your data, build a demand forecast at the right grain and connect it to the inventory or ERP system your planners already work in.

Frequently asked questions

Do I need a demand forecast if I already have a sales forecast?

If you buy, import or manufacture stock, yes. A sales forecast tells finance what to expect in naira; a demand forecast tells purchasing how many units to bring in, at SKU and location level, including demand you missed. Businesses that use the sales forecast for purchasing systematically under-order fast movers.

How many SKUs can this handle?

Modern methods handle thousands of items routinely — a single machine-learning model can be trained across the entire catalogue. The constraint in Nigeria is rarely computing power; it is product master data quality, availability records and consistent unit measures across branches.

What if my business is seasonal and the seasons keep shifting?

Religious calendars shift each year, so supply the dates as explicit features rather than relying on "same week last year". Give the model Ramadan start and end dates, both Eid dates, Easter, school resumption and your own promotion calendar for past and future periods.

Can Excel do demand forecasting?

Excel can run baselines, simple smoothing and scenario tables, and for a business with a few dozen SKUs that may be enough. It becomes impractical with hundreds of items, multiple locations, promotions and hierarchical reconciliation, and it offers no reliable way to track accuracy over time.

How do I forecast when I have only six months of data?

Use a baseline, forecast at family level rather than SKU level, borrow seasonality from similar products or industry patterns, and hold wider safety stock while you accumulate history. Begin recording availability and prices immediately; in six months you will have enough for a proper model.

Who should own the forecast inside the business?

One named planner or purchasing manager, supported by whoever maintains the data. Demand planning fails when it belongs to "IT" or to a consultant. The owner runs the cycle, records overrides, reviews accuracy and decides what changes.

Does AI replace the planner?

No. It removes the arithmetic and the guesswork from hundreds of routine items so that the planner spends their time on exceptions: new listings, supply problems, promotions and the twenty items where judgement genuinely beats the model.

How often should demand forecasts be regenerated?

Weekly for the forecast itself, even when ordering is less frequent, so that emerging shifts are visible early. Retrain the underlying models monthly or quarterly, and rebuild properly after any structural change such as a major price revision or a new distribution channel.

Sources and further reading

Figures, platform rules and regulations change. These are the primary references behind this article and the places to check before you act on it.

• Core Web Vitals — web.dev
• Kubernetes documentation
• Docker documentation
• OpenAI platform documentation
• Anthropic developer documentation