Demand Forecasting for Small Business: When Machine Learning Beats a Spreadsheet

Most small businesses forecast by looking at last year and adding a bit. That works better than people admit, and the honest question is not whether machine learning is more sophisticated but whether it is more accurate for your situation.

Often it is not. Here is how to tell.

What forecasting actually does

A demand forecast predicts how much of something you will need at a point in the future — stock, staff hours, cash.

It works by finding patterns in your history: seasonality, trend, day-of-week effects, and the relationship between demand and external factors like weather, school terms or public holidays.

It does not predict genuinely novel events. No model foresaw the pandemic from prior sales data. Forecasting tells you what happens if the future broadly resembles the past.

When a spreadsheet is enough

Be honest about this before spending anything.

Your demand is stable. If sales vary by less than twenty percent month to month, last year plus a growth factor is accurate enough. A model will improve on it marginally.

You have under two years of data. Seasonality needs at least two full cycles to detect and three to trust. Below that, a model is pattern-matching noise.

You have few product lines. Under about twenty lines, a person who knows the business will usually beat a model, because they hold context the data does not contain.

Your lead times are short. If you can reorder in two days, forecasting matters less. You can react instead of predict.

When machine learning starts to win

Many lines with different patterns. Once you are managing hundreds of SKUs, no human can hold each one's seasonality. This is the clearest case.

Demand responds to external factors. If sales move with weather, school terms, or local events, a model can quantify relationships a person estimates by feel. Weather-sensitive businesses often see the largest improvement.

Long lead times. If you order twelve weeks ahead, forecast error is expensive in both directions. Better accuracy pays directly.

Holding costs are significant. If capital tied up in stock is material, a percentage improvement in accuracy converts straight to cash.

Stockouts cost you customers permanently. Where a stockout means the customer buys elsewhere and does not come back, the asymmetry justifies investment.

What accuracy to expect

Vendors quote accuracy improvements without specifying the baseline, which makes them meaningless.

Realistic figures against a naive last-year baseline:

Note the top row. For many small businesses the honest answer is that a model will not beat what you already do.

Ask any vendor: what is the baseline, and what is the error on data the model has never seen? A model evaluated on the data it was trained on will look excellent and mean nothing.

What you need before starting

Two to three years of clean sales history. By product, by date. Gaps and format changes are the most common reason projects stall.

Records of the things that distort demand. Promotions, price changes, stockouts. A stockout looks like low demand to a model that does not know you had nothing to sell — this single issue causes more bad forecasts than any modelling choice.

A decision the forecast will change. If the number arrives and nobody acts differently, do not build it. This is the most common failure, and it is organisational rather than technical.

What it costs

The ongoing cost is not optional. Models drift as behaviour changes, and an unmaintained forecast quietly becomes worse than the spreadsheet it replaced. Anyone selling a one-off build with no maintenance is selling you a decaying asset.

The test worth running first

Before commissioning anything, take last year's data, produce a forecast using your current method, and compare it against what actually happened. Calculate your average error as a percentage.

Now you have a baseline, and you can ask any vendor to beat a specific number rather than a vague one.

If your current error is already under ten percent, machine learning has little room to help and you should spend the money elsewhere. If it is over thirty percent and the errors are costing you real money, there is a case.

That comparison costs an afternoon and is the single most useful thing you can do before this conversation.