Forecasting is the one quantitative technique that pays for itself in a mid-market company, and it is the one most often abandoned after a pilot. The abandonment is rarely about accuracy. It is about the forecast not being connected to a decision anyone actually makes.

A forecast that nobody acts on is a report. Reports get cancelled in the second budget round.

What follows covers what to settle before modelling, how much history you need, which method to start with, how to tell whether it works, and what a first forecast costs in time for a 40-person distributor or a 150-person consumer-goods brand. None of it needs a data scientist.

Why does forecasting help smaller companies more?

Because smaller companies have less slack to absorb being wrong. A large distributor with regional warehouses can move stock sideways when a forecast misses. A forty-person company with one warehouse cannot. It either holds cash in inventory it does not need, or it disappoints a customer. The cost of error is concentrated, which means the return on reducing error is concentrated too.

This is also why the sophisticated model is rarely the right first move. Most of the available gain sits in the gap between "no forecast" and "a competent simple forecast", not between a competent simple one and a very good one.

What you need before a model is worth building

Three things, in order, and none of them are algorithms.

A clean history of what happened, separated from what you did about it. Sales history is not demand history. If you were out of stock for three weeks, your sales record says demand was zero and your model will learn that. Stockouts, promotions and price changes have to be recorded distinctly, or the model is being taught your past mistakes as if they were customer behaviour.

A decision with a deadline. "How much do we order, and by when" is a decision. "What will next quarter look like" is a conversation. The forecast horizon must match the lead time of the decision. Forecasting six months out when your supplier lead time is three weeks optimises the wrong number.

An agreed measure of wrong. Being 10% over and 10% under are not equally expensive, and most businesses know which one hurts. If you have not written that down, every forecast review becomes an argument about whether the model is good.

Before any model is builtProcess in 4 steps: 1. Name the decision (Order quantity, timing and owner); 2. Separate demand from sales (Log stockouts, promotions, price changes); 3. Agree the cost of error (Over and under are not equal); 4. Build a simple baseline (The benchmark every model must beat).01Name thedecisionOrder quantity,timing andowner02Separatedemand fromsalesLog stockouts,promotions,price changes03Agree thecost oferrorOver and underare not equal04Build asimplebaselineThe benchmarkevery modelmust beatBefore any model is builtProcess in 4 steps: 1. Name the decision (Order quantity, timing and owner); 2. Separate demand from sales (Log stockouts, promotions, price changes); 3. Agree the cost of error (Over and under are not equal); 4. Build a simple baseline (The benchmark every model must beat).01Name the decisionOrder quantity, timing and owner02Separate demand from salesLog stockouts, promotions, price changes03Agree the cost of errorOver and under are not equal04Build a simple baselineThe benchmark every model must beat
Figure 1 Before any model is built. Each step removes a reason the forecast would be ignored. The model comes last, and it has to beat the baseline to earn its place. Source: TechPulse engagement practice

How much history do you need?

For products with a seasonal pattern, two full seasons of clean history is the practical minimum, which usually means two years of weekly or monthly data on stable product codes. One season lets a model see the pattern once; the second lets it separate the pattern from the trend. Steady, non-seasonal lines can start with six to twelve periods of history.

That rule of thumb comes straight from the standard method. The NIST/SEMATECH handbook notes that the Holt-Winters method, the usual choice when demand has both trend and seasonality, needs at least one complete season of data to start, and that two complete seasons are advisable so the trend can be estimated as well.

How strong the seasonal pattern is depends on the category. In the Federal Statistical Office's unadjusted retail turnover index, clothing and footwear rose from 84.4 in February 2024 to 131.6 in December, while food moved from 98.8 to 116.8. A model that treats every line with one seasonal profile will overstock the steady lines and understock the volatile ones.

Swiss retail turnover, February and December 2024Bar chart: Food, February, 98.8; Food, December, 116.8; Clothing, February, 84.4; Clothing, December, 131.6; All retail, February, 93.1; All retail, December, 119.6. Source: [Federal Statistical Office, retail trade turnover](https://www.bfs.admin.ch/bfs/en/home/statistics/catalogues-databases/graphs.gnpdetail.2026-0226.html).Food, February98.8Food, December116.8Clothing, February84.4Clothing, December131.6All retail, February93.1All retail, December119.6Swiss retail turnover, February and December 2024Bar chart: Food, February, 98.8; Food, December, 116.8; Clothing, February, 84.4; Clothing, December, 131.6; All retail, February, 93.1; All retail, December, 119.6. Source: [Federal Statistical Office, retail trade turnover](https://www.bfs.admin.ch/bfs/en/home/statistics/catalogues-databases/graphs.gnpdetail.2026-0226.html).Food, February98.8Food, December116.8Clothing, February84.4Clothing, December131.6All retail, February93.1All retail, December119.6
Figure 2 Swiss retail turnover, February and December 2024. Unadjusted nominal turnover index, average month of 2021 = 100. Clothing swings far more across the year than food, so it needs its own seasonal profile. Source: Federal Statistical Office, retail trade turnover

Stable codes matter as much as length. If a system change renumbered your articles last year, you have one season of history, not three, until someone maps old codes to new. That mapping is the same clerical work that dominates ERP data migration, and it is one of the five foundations in our guide to data readiness for AI.

Which method, and when

MethodWorks whenNeedsTypical first result
Naive or last periodDemand is stable, lead times shortNothingThe baseline everything else must beat
Moving averageNoise dominates, no strong trend6–12 periods of historyBeats intuition, costs nothing
Exponential smoothingClear trend and seasonality2–3 seasons of historyWhere most SMEs should stop
Regression on driversDemand tracks a known external factorThe driver data, historicallyUseful when drivers are real
Machine learningMany SKUs, long history, real seasonalityYears of clean data, ongoing ownershipRarely first, sometimes right

Exponential smoothing deserves its place in the middle of the table. It weights recent observations more heavily than older ones, comes in single, double and triple forms for level, trend and seasonality, and is built into spreadsheets and most ERP planning modules. Plain averages, by contrast, are only a sensible forecast when there is no trend.

The evidence favours starting simple. In a peer-reviewed comparison on 1,045 monthly series from the M3 forecasting competition, established statistical methods beat the machine learning methods tested on both accuracy measures and at every forecast horizon, with far lower computing requirements. Machine learning earns its place when there are thousands of related items and years of clean data, which is the territory of a national retailer more than of a 60-person wholesaler.

The last row is the one companies ask about first. It is legitimate, but it is also the only row that creates a permanent maintenance obligation, and the row most likely to be abandoned when the person who built it leaves. When that obligation is worth taking on is discussed in getting AI past the pilot stage.

Do we need machine learning or a data scientist?

Usually not. A company of 20 to 300 people can run exponential smoothing in a spreadsheet, in the planning module of its ERP or in an inexpensive planning tool. What it needs is an analyst who understands the business and a named owner for the forecast. A data scientist becomes worthwhile when there are thousands of items, several real demand drivers and a forecast feeding automated orders.

Missing expertise is the most common reason companies give for not using AI at all. Across the EU in 2025, 71% of enterprises that had considered AI but not used it cited a lack of relevant expertise, according to Eurostat. Forecasting is a place where that barrier is lower than it looks, because the methods that work first are decades old, publicly documented and already inside ordinary tools. The scarce resource is not modelling skill. It is someone who knows why the March numbers were odd.

Public statistical series can be genuine drivers rather than decoration. The Federal Statistical Office publishes monthly retail turnover, calendar and seasonally adjusted, and SECO publishes a consumer sentiment index. National series rarely move a single product, but at category level we have seen one such series explain more than months of model tuning. Where the drivers are many and the items number in the thousands, our AI and machine learning work picks up where this article stops.

How do you know the forecast is working?

Not by the accuracy metric alone, but by whether the decision changed and the outcome improved. Run the forecast in parallel with current judgement for one full cycle without letting it drive orders, then compare both against what actually happened, weighing the cost of each error rather than its size. The comparison shows where the model should be used and where people should be left alone.

Judgement often wins on the fast-moving lines that buyers watch every day, and loses badly on the long tail nobody has time for. That tells you exactly where to deploy the model. Textbook measures such as mean squared error, which the NIST handbook uses to compare methods, treat a shortfall and an excess alike. Your costs do not, which is why the agreed measure of wrong from the first section belongs in the comparison.

This parallel run is also what makes adoption possible. A forecast imposed on a planner who was not consulted gets overridden quietly, and an overridden forecast produces no benefit while still costing maintenance. Record overrides and their reasons. After a season, they are the best training material you have.

A worked example: a 40-person distributor

Take a 40-person distributor of household goods with a few thousand active items, one warehouse, and suppliers in Europe and Asia with lead times of three to ten weeks. The buyer orders from experience, and the finance lead worries about stock value.

Rank before you model. In the distributors we have worked with, a few dozen lines usually account for a large share of the cash tied up in stock. Rank items by average stock value and by how much their demand varies, and place them on the matrix below.

Where to point the first forecastTwo-by-two matrix. Horizontal axis: Demand variability. Vertical axis: Cash tied up in stock. Top left: Steady big sellers, Simple method, review monthly. Top right: Volatile cash tiers, Forecast first, review weekly. Bottom left: Steady small lines, Reorder points, no model. Bottom right: Erratic long tail, Judgement and safety stock.Steady big sellersSimple method, review monthlyVolatile cash tiersForecast first, review weeklySteady small linesReorder points, no modelErratic long tailJudgement and safety stockCash tied up in stockDemand variabilityWhere to point the first forecastTwo-by-two matrix. Horizontal axis: Demand variability. Vertical axis: Cash tied up in stock. Top left: Steady big sellers, Simple method, review monthly. Top right: Volatile cash tiers, Forecast first, review weekly. Bottom left: Steady small lines, Reorder points, no model. Bottom right: Erratic long tail, Judgement and safety stock.Steady bigsellersSimple method,review monthlyVolatile cashtiersForecast first,review weeklySteady smalllinesReorder points, nomodelErratic longtailJudgement andsafety stockCash tied up in stockDemand variability
Figure 3 Where to point the first forecast. Start where cash and volatility meet. Steady lines need reorder points, and the erratic long tail needs safety stock rather than a model. Source: TechPulse engagement practice

Clean twenty lines, not three thousand. Take the twenty items in the top-right quadrant. Mark the weeks they were out of stock and the weeks they were on promotion, and map any renumbered codes. That is days of work, not months.

Forecast to the lead time. Fit seasonal exponential smoothing at monthly level, with a horizon equal to the supplier lead time plus the review period. Compare it with a naive seasonal baseline, last year's same month, and keep the model only where it wins.

Run one season in parallel. The buyer keeps ordering. At the end of the season, price both sets of errors: excess stock at its carrying cost, shortfalls at the lost margin. Keep the model for the lines where it was cheaper, and leave the rest to the buyer.

What does a first forecast cost, and how long does it take?

Mostly internal time. For a company of 20 to 300 people, a first forecast for the twenty to fifty lines that matter most typically takes four to eight weeks of part-time work to build, followed by one full cycle of parallel running before it drives orders. Software is rarely the constraint, because spreadsheets and existing ERP modules handle the methods that work first.

These are typical observations from the engagements we have run, not a quote, and the spread depends mostly on how clean the history is. What changes with size is governance rather than method. At 20 people the owner is usually the buyer, and the forecast lives in a spreadsheet reviewed weekly. At 200 people there are several buyers, the forecast should live in the ERP, and someone has to decide who may override it. If you are not yet sure your system can carry that, see when a small company actually needs an ERP.

Measure the result in the currency the owners care about: stock value, days of inventory and the share of orders delivered complete. The structure for putting that case before you spend is in how to measure ROI on a digital transformation.

Where this sits in a wider programme

Forecasting touches inventory, purchasing and finance at once, which makes it a useful first quantitative project and a poor first system project. If the underlying stock records are unreliable, fix those first. A forecast built on inventory data that does not match the shelf will be blamed for a problem it did not cause. Our quantitative strategies practice starts with a short data assessment before any modelling for exactly this reason, and it regularly ends with a recommendation to postpone the model.

For retail and consumer goods businesses, the highest-value first cut is usually not all items at once. It is the twenty lines that tie up the most cash, forecast properly and reviewed weekly by a named person. That is measurable by the following quarter, and it belongs on the one-page digital strategy as one of the few changes the company commits to.

Sources

Working on this right now?

Tell us where you are in two questions. A consulting partner reviews every enquiry within 2 business days.