Sales Forecasting With Machine Learning for SMEs

The spreadsheet forecast has a hidden cost

In most small and mid-sized businesses, the sales forecast is a spreadsheet: last year's number, plus a growth percentage the managing director feels good about, adjusted for anything obvious. It is quick and everyone understands it.

The cost shows up elsewhere. Stock ordered against an optimistic forecast sits in the warehouse. Recruitment plans slip when the quarter comes in low. Worst of all, the forecast is built by the same people whose targets depend on it, so it tends to say what they hope rather than what the data suggests.

What machine learning adds, and what it does not

A machine learning forecast is still built from history. What it adds is the ability to account for many drivers at once and to do it consistently every week for every product, region or customer group.

It does not predict events that have never happened. A new competitor, a lost key account or a supply shock will break every method equally. The honest benefit is a better baseline and a clear view of uncertainty, which leaves managers free to argue about the genuinely unknowable parts. Our post on forecasting realistically goes further into what to expect.

How much data an SME actually needs

For weekly or monthly sales with a yearly cycle, you want at least two full years, preferably three, so the model sees each season more than once. More important than length is consistency: product codes that did not change meaning halfway through, and records of past promotions and price changes.

• Sales by week or month at the level you plan at (product family, branch, channel)
• Promotion and discount calendars, even if they live in someone's inbox
• Price changes and stock-outs, because a month of zero sales due to no stock is not zero demand
• For B2B, the CRM pipeline with stage history
• Known calendar effects: bank holidays, school terms, Ramadan and Eid for Gulf businesses, trade shows

The stock-out point matters more than people expect. A forecast trained on sales rather than demand learns that you sell nothing when you run out, and keeps recommending you order too little.

A realistic SME example

Consider a 60-person catering equipment distributor with 2,000 products, sales to restaurants and hotels, and a sales forecast built by product manager spreadsheet. Top sellers are forecast reasonably well. The long tail is guessed, and the business carries a lot of slow stock as a result.

A model forecasting at product-family level, using price changes, promotions, season and the hospitality calendar, could reasonably reduce forecast error on the mid and long tail, where human attention never reaches. The top 50 products might barely improve, since people already watch them closely. That split is normal and worth knowing before you budget: the gains from machine learning sales forecasting come from breadth, not from beating an expert on the lines they know best.

Judging a forecast honestly

Before anyone trusts a new forecast, test it the way it will be used: train on data up to a past date, forecast the following months, and compare with what really happened. Repeat for several cut-off dates.

1. Pick an error measure the business understands, such as average percentage error weighted by revenue
2. Compare against a naive forecast (same as last year) and your current spreadsheet method
3. Look at bias separately from error: consistently high forecasts hurt stock more than noisy ones
4. Report a range around every number, and check that reality falls inside them as often as claimed

A forecast that is 8% more accurate but that nobody uses is worth less than the spreadsheet everyone already reads.

When it is the wrong project

If sales are dominated by a handful of large, lumpy contracts, a forecast model has little to learn from and a pipeline review will do better. The same applies to businesses under two years old or those that have just changed what they sell.

And if the forecast does not change any decision, including stock, staffing or cash, the accuracy improvement has no value. We ask that question first at SpiderHunts, before looking at any data. For businesses whose main question is stock rather than revenue, our post on time series forecasting for demand planning is closer to the mark.

What building one involves

A first machine learning sales forecast for an SME is typically a six to ten week piece of work: data cleaning, back-testing against the current method, a forecast refreshed weekly or monthly, and output delivered where planners already work, usually a spreadsheet or BI tool rather than a new app. Our data science team keeps the first version deliberately plain so it can be explained in a board meeting.

The ongoing cost is small, but there must be an owner who checks accuracy each month and notices when the business has changed enough that the model needs retraining.

Frequently asked questions