When to Trust Your Forecast (and When to Ignore It)

Adjusting a system forecast before placing an order is common practice. Across four supply chain companies, Robert Fildes and Paul Goodwin found an average of 75% of statistical forecasts were adjusted before use.

Whether those adjustments make the forecast better is something you can test, and so is the accuracy of the forecast before anyone touches it. Both start from the same two numbers.

‍

Score the Forecast You Already Have

To score a forecast you need two numbers for each product: what you forecast, and what you sold.

You already have the sales figures. Start saving the forecasts too, so you have something to compare them against.

Then work out two measures, because they tell you different things.

• How big the misses are. For each product in a period, take the difference between forecast and actual, ignore whether it is over or under, and turn it into a percentage of what actually sold. Average those across a category and you have what is usually called mean absolute percentage error, or MAPE. It tells you how much slack to build into a buying decision.
• Which direction the misses run. Do the same calculation again, but keep the plus and minus signs so overs and unders cancel out. If the result is close to zero, your errors fall on both sides. If it sits high or low every time, you have bias. Bias usually has a specific cause: returns not netted off, promotions missing from the sales history, or a lead time that changed after you set the forecast up.

Work these out by category rather than across the whole range, or your steady lines will average out the volatile ones.

‍

Why Some Products Are Harder to Forecast

How predictable a product is affects accuracy as much as the method you use to forecast it.

• Steady repeat lines sell at a fairly consistent rate week to week, so history is a reasonable guide, and you can usually follow the forecast without much adjustment.
• Seasonal lines have a shape that repeats, so they forecast more reliably once you have two or three previous seasons to work from, and less so in year one.
• Promotion-driven lines can look volatile when you measure promotional weeks and full-price weeks together. Separate them and the underlying rate is often steadier.
• New products and trend-led lines have no history behind them, so the forecast is closer to an assumption than a projection.

One rule across the whole range is unlikely to fit all of it, so set the approach group by group.

‍

Sort Your Range by Predictability

Take your top lines by revenue and put each into one of three groups, using the accuracy figures above rather than your impression of them.

• Predictable. Small average error, little bias. Follow the forecast and hold modest cover.
• Variable but understood. Larger errors, but with a pattern you recognise, such as seasonal peaks or promotional spikes. Follow the forecast, hold more cover, and review before each peak.
• Unpredictable. Large errors with no pattern, or no history at all. The forecast carries limited information here, so order smaller quantities more often and keep the option to reorder.

The first group needs the least attention, which leaves more time for the groups where your judgement is doing the work.

‍

When Overriding Makes Sense

A forecast built from sales history reflects what has already happened. Adjust it when you know something it does not:

• A promotion, discount or bundle that is planned but has not run before.
• A product launch, a relaunch, or a line you are about to delist.
• A supply disruption, a lead time change, or a supplier minimum that forces a different order size.
• A new sales channel going live, which adds demand with no history behind it.
• A one-off event such as a press feature, a collaboration, or a competitor leaving the category.

In each case, you are adding information the forecast did not have, rather than disagreeing with it.

Each time you do, record three things: the figure the system gave you, the figure you went with, and the reason in a few words. After several cycles, score both figures against actual sales using the measures above. 

That shows you where your adjustments improved on the forecast and where they did not. Fildes and Goodwin found that small adjustments made little difference to accuracy and in places reduced it, so the comparison is worth running before adjusting becomes routine.

‍

Tracking Accuracy Without Building a Spreadsheet for It

You can do all of this by hand, and it competes for attention with everything else in a buying cycle.

Having it in the system you already buy from takes the manual step out. Brightpearl by Sage is a Retail-First ERP System, which replaces min and max spreadsheets with seasonal demand forecasting and suggested purchase orders. 

Sales, stock and purchasing sit in one system, so forecast and actual land against the same products, and you can see stockouts, slow movers and sell-through by channel and location.

Kirstin Smith, Chief Operations Officer at Miron Crosby, describes the balance this is aiming at: 

"With Brightpearl we're not over-ordering, but we're also not missing opportunities."

A forecast does not have to be accurate to be useful. It has to be measured, so that following it or overriding it is a decision you can back.

‍

Ready to plan purchasing on numbers you can check?

Discover how Brightpearl brings forecasting, purchasing and live stock into one system. Book a demo today.