Forecasting and demand planning on the data you already have

Most forecasting projects stall not on the method but on a misbelief that the data is not good enough yet. In practice, a straightforward trend model on clean historical data outperforms a sophisticated model on data nobody trusts. And you already have the historical data.

Almost every business forecasts. Few do it well. The usual method is a spreadsheet, last year’s number, a bit of gut feel, and a hopeful adjustment for growth. It is better than nothing and a long way short of what the data you already keep would allow. A proper forecast is not a new system or a new pile of data. It is a better use of the order history sitting in the systems you already run on.

You already forecast. Just not well.

The spreadsheet forecast has three weaknesses, and they are predictable. It leans on one person’s judgement, so it carries their blind spots. It tends to produce a single number, which is always wrong by some amount and tells you nothing about how wrong it might be. And it only ever covers the handful of lines someone has time to think about, while the long tail is guessed or ignored. None of that is a failure of effort. It is the limit of doing it by hand.

What a real forecast adds.

A proper forecast reads the history and pulls out what the eye misses. The seasonality, the lines that always lift before a particular time of year. The underlying trend, separated from the noise. And crucially, a range rather than a single figure, a sense of the likely high and low, which is what lets you plan stock and staffing against risk rather than against a single hopeful guess. It also scales, producing a forecast for every line rather than the few you could manage manually, so the long tail stops being a blind spot.

It runs on the sales history you keep.

The reassuring part is that this needs no new data. Your order and sales history, the records you keep to invoice and account, are exactly what a forecast learns from. Two or three years of it is usually plenty for a mid-market business. You have been collecting the training set the whole time. The work is to read it properly, not to go and gather something new.

Where forecasting pays, and where it can’t.

Be honest about the limits. Forecasting pays where demand has some pattern to it, seasonality, trend, the rhythms of how your customers buy, which covers most established products and most of the order book. It cannot conjure a pattern where none exists. A brand new product with no history, a one-off, or a market about to be upended by something the data has never seen, are beyond it, and a forecast that pretends otherwise is worse than none. Use it where the past is a fair guide to the future, and use judgement where it is not.

Start with one line that matters.

You do not forecast the whole catalogue on day one. You take one product or category that matters, where being wrong is expensive, and you forecast that properly first. It proves the approach on something that counts, it is quick, and it gives you a real comparison against the way you forecast it today. The second line is faster, and the method spreads once people see it beat the spreadsheet.

So this week, take your most important product or category, find the forecast you made for it last year by eye, and lay it against what happened. The gap is the cost of forecasting by hand, and it is usually larger than people expect. Building forecasts on the history you already hold, and surfacing them in the reports you already use, is everyday machine learning and Power BI work for us, and it is the natural next step once your reporting is trusted.

The data to do it properly is already yours.