Stop trying to improve your forecast accuracy.

Almost every planning improvement program I've seen starts the same way. Someone measures forecast error, discovers it's worse than they hoped, sets a target to reduce it, and buys or builds something to help. Two years later the error is roughly where it was, and the honest conclusion drawn is usually "our business is just hard to forecast."

Sometimes that's true. More often the program aimed at the wrong variable. Forecast accuracy is a real thing worth measuring, but in most mid-market companies it is not the binding constraint on planning performance — and improving it wouldn't help much even if you could.

The number of hands is the variable

Here's the thing nobody measures: how many people are permitted to change the number between the system producing it and the purchase order going out.

In a typical distributor, the answer is something like five. The statistical forecast produces a quantity. The planner adjusts it based on judgment. A sales rep who knows the account pushes for more. The vendor rep mentions a price break at a higher quantity. Someone senior looks at last year and rounds. Each of those interventions is individually defensible, made by a competent person acting in good faith with real information.

The aggregate is chaos. Five sequential adjustments to one number, made in different rooms, on different information, with different incentives, none of them recorded. Improving the accuracy of step one — the statistical forecast — by five points does approximately nothing when steps two through five can move it by 300%.

Why this is the harder problem, and therefore the skipped one

Forecast accuracy is a comfortable problem. It's technical, it belongs to one team, it has a metric, and you can buy something that claims to address it. Nobody's authority is threatened by an improved algorithm.

Decision rights are the opposite. Deciding that a sales rep may flag but not change a number is a political act. Telling a senior person their instinct now goes through a documented override rather than a phone call is a political act. This is why the technical fix gets funded and the governance fix gets postponed, and it's why so many planning transformations produce better inputs to an unchanged process.

It's also why this work needs someone with operating authority behind it, or a sponsor willing to lend theirs. A consultant with a good algorithm can improve your forecast. Nobody can fix your decision rights from outside the room.

What to do instead

1. Count the hands, then draw the line

Map every point where a number can change between system output and purchase order. Write down who can change it, on what basis, and whether anyone is notified. Most companies have never drawn this map, and drawing it is often enough to embarrass the process into changing.

Then draw the line: who decides, versus who informs. A sales rep who knows an account is about to place a large order has genuinely valuable information — and it should enter the process as an input to the demand review, not as an edit to a quantity on a Thursday afternoon. The distinction between "you may tell us" and "you may change it" is the whole ballgame.

2. Make overrides cheap to record and expensive to hide

The goal is not to eliminate human judgment. Judgment is often right, and a planning system that can't absorb it is worse than one that can. The goal is that the override is written down with its reasoning, and then checked later against what actually happened.

Do that for two quarters and you learn something no forecasting tool will tell you: which of your people are adding value with their adjustments and which are adding noise. In my experience it splits about evenly, and it is never the split anyone predicted. That finding alone is worth more than a five-point accuracy improvement, because it tells you where to expand judgment and where to constrain it.

3. Define the number once

You cannot arbitrate between two people who disagree about the forecast if you can't agree on whether last month was good. This is why the definitional fight comes before the process fight, and the process fight comes before the technology. Companies that skip to step three end up automating an argument.

The honest counterargument

There are businesses where forecast accuracy genuinely is the constraint — where the process is disciplined, the decision rights are clean, the overrides are recorded, and the residual error is real and expensive. High-volume, short-shelf-life, promotion-heavy businesses often live here. If that's you, buy the good forecasting engine; you'll get the return. I've led three of those implementations, so this isn't me talking you out of software.

But that's a company with its planning discipline already in place, and it's a small minority of the mid-market. The test is easy: ask three people in your business what your fill rate was last month. If you get three answers, you don't have a forecasting problem yet. You have a prerequisite.