A statistical baseline gives you a good-enough forecast for most of your portfolio, so people can spend their time on the few products that drive most of the value. Clean the history with simple rules first, then let the system test four or five complementary methods per product and keep the best one. Just starting out? A weighted moving average already gets you a long way.
Watch the 7-minute video this article is based on, or see the full S&OP training series.
Why start with statistics at all?
Most businesses make most of their money from a handful of products: the flagship items, the cash cows. The forecasting game is really about one question: how do we spend as much time as possible on the products that create the most value?
That means not getting pulled into hundreds of B and C items that don't move the needle. Statistics is one of the best tools for this, because it gives you a forecast that is good enough for the large majority of items. Good enough here means you can hit your target service level without blowing up inventory with safety stock to buffer every unknown.
The two things that matter
There are dozens of forecasting methods. You only need to be sharp on two questions:
- How do you consistently clean your sales history, so noise that shouldn't shape the future doesn't make the forecast worse?
- Which method or mix of methods gives the best forecast performance across your whole portfolio?
1. Cleaning the sales history
Don't worry too much when you start out. It won't be perfect, and simple rules work. A common one: exclude anything more than two standard deviations from the norm.
You will sometimes remove a real signal that way. As you mature, document which peaks were genuine and add rules that keep them in. Your forecast improves with every cycle you do this.
2. Choosing forecasting methods
All forecasting methods try to do the same thing: predict the future from patterns in the past. But ice cream doesn't sell like winter tyres, so some methods work better on some products than others.
The best overall result comes from a mix: four or five methods that your computer or planning platform runs on every product, keeping whichever performs best for each one.
A method mix that works
Pick methods that complement each other. A combination I have found works well:
- Holt-Winters looks for a baseline, a trend and seasonality.
- ARIMA looks for the same three things and adds autocorrelation, picking up other patterns in the history.
- SBA (Syntetos-Boylan approximation) handles intermittent demand: orders that arrive in chunks and are hard to predict.
- Simple moving average is the catch-all. When the other three don't find a clear pattern, it usually gives the best result.
Just starting out? Use a weighted moving average
Running several models per product is a tall order if your S&OP journey has just begun. A simple moving average already gives a decent result. To sharpen it a little, weight the recent past more heavily, for example:
Forecast = 0.5 × average of the last 6 months + 0.5 × average of the last 3 months
It is still simple, but it reacts faster to recent changes and nudges accuracy up. Moving to automatic method selection per product will give you a substantially better result across the portfolio once you are ready for it.
With the statistical baseline in place, the next step is to enrich it with what the sales teams know. That starts with deciding which affiliates and markets to involve.
Want to see this implemented for your team?
OmniaOps is a supply chain planning platform built by ops practitioners. 30-minute call, no slide deck.
Book a 30-min call