Is a more accurate forecast always better?

No. A forecast has no value on its own, its value is borrowed entirely from the decision it informs. The right measure is not how close the forecast lands to the eventual actual, it is whether it lands on the correct side of the decision it drives. Two forecasts with identical error can have wildly different value: one sitting comfortably far from any decision threshold can be sloppy and cost nothing, while one sitting right on a threshold must be sharp because a small error flips the action. Improving a forecast that is already on the right side of every threshold it feeds produces a nicer number and no better outcome.

How do I know whether improving a forecast is worth the effort?

Run it through two questions. First, what decision does this forecast drive, and where is the threshold at which that decision changes? Build the line if demand exceeds 10,000 units, add a shift if volume passes a capacity limit, reorder when stock falls below a reorder point. Second, how far is the forecast from that threshold relative to its typical error? Divide the distance to the threshold by the typical error. Comfortably above 1 and the decision is safe, so accuracy buys nothing. At or below 1 and plausible error could flip the call, so accuracy is decisive and worth every hour.

Why did reducing forecast error from 12% to 7% change nothing?

Usually for one of two reasons. Decisions often round away precision: you order in full pallets, staff in whole shifts, and build in discrete lines, so precision below the size of one lump is wasted, and both answers round to the same 9 pallets. Or a buffer already absorbs the miss: safety stock, contingency and slack exist precisely to make decisions robust to forecast error, so where the buffer already covers the plausible error, tightening the forecast underneath it improves a number the buffer has already made irrelevant. In both cases the forecast got better and the decision did not move.

Should I report MAPE to leadership, or something else?

Report proximity alongside error, not error alone. A decimal of MAPE tells a CFO how the forecasting process performed. A statement such as comfortably above the build threshold even in the low case tells them what to do, which is what they are actually deciding. The related habit worth building is triage: rank your forecasts by distance to the thresholds they feed, spend analytical time on the ones sitting near a tipping point, and stop apologising for error on forecasts that drive no marginal decision.

When does the decision-relevance lens not apply?

It fits a forecast tied to specific decisions with identifiable thresholds. It applies less cleanly when one forecast feeds many small decisions whose errors aggregate, such as cash forecasting, where lots of modest misses can compound into a real total. In that situation accuracy matters more broadly because the aggregate is itself the decision. The test is not never improve a forecast. It is know which decision your improvement is supposed to change, and confirm that it actually would.

Stop Optimizing Your Forecast Accuracy

Core Argument

An FP&A team spent 3 weeks rebuilding its demand forecast. Mean error fell from about 12% to 7%. It was a real improvement, and the team presented it proudly. When someone asked what the company would now do differently, the room went quiet, because the honest answer was: nothing. The production plan, the staffing, and the purchasing came out identical at 12% error and at 7%. A forecast exists to support a decision, and accuracy is only worth improving when it would change what you decide.

The core insight

A forecast has no value on its own. Its value is entirely borrowed from the decision it informs.

That means the right measure of a forecast is not how close it lands to the eventual actual. It is whether it lands on the correct side of the decision it drives. Two forecasts with the same error can have wildly different value: one that sits comfortably far from any decision threshold could be sloppy and cost nothing, while one that sits right on a threshold must be sharp, because a small error flips the action.

Accuracy, in other words, only matters near the point where a decision changes. Far from that point, precision is a luxury. Near it, precision is everything. Effort spent improving a forecast that is already on the right side of every threshold it feeds is effort that produces a nicer number and no better outcome.

"A forecast is not trying to be correct. It is trying to be correct enough to make the decision it feeds come out right."

The framework: the decision threshold

Before improving any forecast, run it through two questions.

1

What decision does this forecast drive, and where is the threshold?
Almost every forecast feeds a decision with a tipping point. Build the new line if annual demand exceeds 10,000 units. Add a shift if volume passes a capacity limit. Reorder when projected stock falls below a reorder point. Find that threshold. It is the only number the forecast is really being measured against.
2

How far is the forecast from the threshold, relative to its uncertainty?
If the forecast sits far from the threshold compared to how wrong it might be, the decision is safe and more accuracy buys nothing. If it sits close, so that plausible error could push it across, accuracy is decisive and worth every hour.

The test in one line

Headroom = Distance to Threshold ÷ Typical Forecast Error

Comfortably above 1, the decision is safe and accuracy buys nothing. At or below 1, plausible error flips the call and accuracy is decisive.

Two cases make the pattern concrete

Suppose the decision is to build a new production line if forecast annual demand exceeds 10,000 units.

Line	Case A	Case B
Decision threshold	10,000 units	10,000 units
Forecast	14,000 units	10,300 units
Typical forecast error	±1,000 units	±1,000 units
Headroom	4.0	0.3
Could error flip the decision?	No	Yes
Value of more accuracy	Near zero	High

In Case A, the forecast is 4,000 units clear of the threshold and the error is 1,000. No realistic miss changes the answer: you build the line. Refining that forecast from 14,000 to a "better" 13,500 is analytically satisfying and commercially pointless. In Case B, the forecast sits 300 units above a 10,000 threshold with a 1,000-unit error band. Here the decision genuinely hangs on the forecast, and every point of accuracy you can buy directly reduces the risk of a wrong, expensive call. Same threshold, same error, opposite conclusions about where to spend effort.

Why experienced managers get this wrong

1

They treat accuracy as an end in itself.
Forecast error is easy to measure and improving it feels like unambiguous progress. It becomes a target in its own right, disconnected from whether any decision moves. Better MAPE becomes the goal instead of a means.
2

They spend effort uniformly instead of where decisions are sensitive.
Most teams try to improve the whole forecast evenly. The payoff is not even. It concentrates entirely around the thresholds. A forecast far from every tipping point deserves almost none of your attention no matter how large its error.
3

They forget that decisions often round away precision.
Many decisions come in lumps: you order in full pallets, staff in whole shifts, build in discrete lines. When the decision is granular, forecast precision below the size of one lump is wasted. Cutting error from 12% to 7% changes nothing if both answers round to the same 9 pallets.
4

They ignore that buffers already absorb small errors.
Safety stock, contingency, and slack exist precisely to make decisions robust to forecast error. Where a buffer already covers the plausible miss, tightening the forecast underneath it improves a number that the buffer has already made irrelevant.

Putting it to work tomorrow

Start from the decision, not the forecast. For each forecast you produce, write down the decision it drives and the threshold at which that decision changes. If you cannot name one, ask why the forecast exists at all. That question alone often retires forecasts nobody uses.
Triage your forecasts by distance to threshold. The headroom calculator runs this test one forecast at a time. The ones sitting near a decision tipping point are where accuracy pays. The ones sitting comfortably clear are where "good enough" is genuinely good enough. Spend your analytical time accordingly, and stop apologizing for the error on forecasts that drive no marginal decision.
Report proximity, not just error. Alongside the forecast, note how close it is to the decisions it feeds. "Comfortably above the build threshold even in the low case" is more useful to a CFO than a decimal of MAPE, because it speaks to the decision instead of the metric.

Know the limits

Decision-relevance is the right lens for a forecast tied to specific decisions with identifiable thresholds. It applies less cleanly when one forecast feeds many small decisions whose errors aggregate, such as cash forecasting, where lots of modest misses can compound into a real total. There, accuracy matters more broadly and the aggregate is the decision. The test is not "never improve a forecast." It is "know which decision your improvement is supposed to change, and confirm it actually would."

The principle worth keeping

A forecast is not trying to be correct. It is trying to be correct enough to make the decision it feeds come out right.

"Before you spend a day sharpening a forecast, find the decision it drives and how close it sits to changing. If the decision would not move, neither should your effort."

Free Excel Model: Forecast Scenario & Sensitivity

A driver-based forecast with explicit scenario ranges and a sensitivity ranking, so you can see which assumptions move the outcome across a decision threshold and which ones never will.

Download Free ↓