The Hidden Bias That’s Sabotaging Your Demand Forecasts

Lately LinkedIn is buzzing with ML experts making outsized promises that automated algorithms crush human forecasts, and planners shouldn’t touch a model ever again.

The favorite talking point? 👇

 𝗔 𝘀𝗶𝗺𝗽𝗹𝗲 𝗺𝗼𝘃𝗶𝗻𝗴 𝗮𝘃𝗲𝗿𝗮𝗴𝗲 𝗯𝗲𝗮𝘁𝘀 𝘆𝗼𝘂𝗿 𝗵𝘂𝗺𝗮𝗻 𝗳𝗼𝗿𝗲𝗰𝗮𝘀𝘁.

Sounds alarming. Look closer, and it falls apart. It’s not the inability of the planner, it’s the bias in the planning process that just lifted the forecast up.

S&OP/IBP, what have you — the powers that be said the forecast needs to come up 25% across the board.

A clean, highly seasonal pattern just lifted up magically across the entire horizon – right seasonality, right peaks and troughs.

But accuracy comes back at zero. 📊

𝗧𝗵𝗮𝘁 𝗯𝗶𝗮𝘀 𝗱𝗶𝗱𝗻’𝘁 𝗰𝗼𝗺𝗲 𝗳𝗿𝗼𝗺 𝘁𝗵𝗲 𝗺𝗮𝘁𝗵 𝗼𝗿 𝘁𝗵𝗲 𝗺𝗼𝗱𝗲𝗹 – it was crafted intelligently by the demand planner.

Now the 6-month moving average proposed by a simple ML algorithm beats this forecast, a genuinely dumb forecast, but unbiased.

It cuts right through the middle of the data, 𝘀𝗼 𝗶𝘁 𝘀𝗰𝗼𝗿𝗲𝘀 𝗽𝗼𝘀𝗶𝘁𝗶𝘃𝗲 𝗮𝗰𝗰𝘂𝗿𝗮𝗰𝘆 𝗮𝗻𝗱 𝘁𝗵𝗲 𝘀𝗺𝗮𝗿𝘁 𝗳𝗼𝗿𝗲𝗰𝗮𝘀𝘁 𝗽𝗼𝘀𝘁𝘀 𝗻𝗲𝗴𝗮𝘁𝗶𝘃𝗲 𝗙𝗩𝗔. 🎯

Fix the bias, and a plain Holt-Winters model hits ~99% accuracy on that same pattern.

The moving average? A consolation prize at best. ✅

The first thing to diagnose when you assess an IBP process is whether a process bias is polluting the demand forecast.

💬 “𝗪𝗵𝗲𝗿𝗲 𝗱𝗼𝗲𝘀 𝗯𝗶𝗮𝘀 𝗰𝗿𝗲𝗲𝗽 𝗶𝗻𝘁𝗼 𝘆𝗼𝘂𝗿 𝗳𝗼𝗿𝗲𝗰𝗮𝘀𝘁 𝗽𝗿𝗼𝗰𝗲𝘀𝘀 – 𝘁𝗵𝗲 𝗺𝗼𝗱𝗲𝗹, 𝗼𝗿 𝘁𝗵𝗲 𝗰𝗼𝗻𝘀𝗲𝗻𝘀𝘂𝘀 𝗺𝗲𝗲𝘁𝗶𝗻𝗴?”

Read our related posts:

• How moving average can beat Machine Learning?
• Machine Learning (ML): Is it still a Buzzword?
• S&OP is not consensus – It is conflict resolution!