Forecast confidence explained

Forecast confidence is the degree of certainty around a revenue forecast, usually expressed as a range: "we expect EUR 2.0 million, and with 80 percent confidence it will fall between EUR 1.9 and 2.15 million." You calculate that range from how far your previous forecasts were off, or by adding up the uncertainty per deal and per customer. A range is only worth something if it holds: at 80 percent confidence, the outcome should fall within the range roughly eight times out of ten.

Why one number is not enough

A forecast of EUR 2,000,000 says nothing about how certain it is. It may be a figure that has been within two percent for four quarters running, or one that depends on a single EUR 400,000 deal that may or may not be signed next week. For the board, the bank or the shareholder, that difference matters more than the figure itself.

A range also forces a more honest conversation. Instead of "will we make it?", the question becomes: "what would have to happen for us to land at the bottom of the range, and what do we plan to do then?" That is a question you can manage against.

Three ways to set a range

1. From your own historical errors

The simplest and often the most honest method. You look at how far your forecasts were off in the past, and use that spread as your margin.

Worked example: suppose you have made a forecast at the start of each of the last eight quarters. The deviations from the actual outcome were, from low to high: -9, -4, -2, +1, +3, +5, +6 and +12 percent. The middle six of the eight lie between -4 and +6 percent. For your new forecast of EUR 2,000,000, a range of roughly 75 percent confidence is then EUR 1,920,000 to EUR 2,120,000.

This method needs no model and works for any forecasting method, including a spreadsheet. The condition is that you keep your forecasts. If you have never done so, start now.

The same approach works for the base forecast from invoices, as described in forecasting from historical revenue.

2. Bottom-up, per deal and per customer

Every open deal has a probability. Every recurring customer has a probability of staying, growing or shrinking. If you have those probabilities, you can calculate how the total outcome is distributed.

The key insight: the expected value is often an outcome that cannot occur.

Worked example: suppose you have a fairly certain base of EUR 1,200,000 and three large deals of EUR 300,000 each, each with a 50 percent chance and independent of one another. The weighted forecast is EUR 1,200,000 + 3 x EUR 150,000 = EUR 1,650,000. But the possible outcomes are:

EUR 1,650,000 does not occur. The realistic outcome lies between EUR 1,500,000 and EUR 1,800,000, with a one in eight chance of EUR 1,200,000. That chance of the bottom outcome is exactly what a board wants to know, and what a weighted total hides.

With many deals, systems calculate this with a simulation: drawing thousands of times which deals are won, and looking at how the totals are distributed. This is called a Monte Carlo simulation. The result is a range that accounts for concentration: a pipeline with three large deals gives a much wider margin than a pipeline of thirty small ones with the same total value.

This method is only as good as the probabilities per deal. If those are wrong, for example because old deals still sit at 50 percent, the range is as false as the figure. Filter first, as in how do you spot an unreliable pipeline.

3. With scenarios

Instead of a statistical margin you choose explicit scenarios: what if the largest customer cancels, what if the two large deals slip a quarter, what if indexation comes out lower. That gives no percentage, but it does give a bottom and a top with a story attached. The approach is worked out in scenario forecasting for B2B.

In practice you combine them: a statistical range for normal variation, and a few scenarios for the events that fall outside it.

How do you check whether your forecast range is right?

A range is a promise. An 80 percent range promises that the outcome falls within it in roughly eight out of ten periods. You can check that. This is called calibration.

1. Keep every forecast with its range.
2. After each period, note whether the outcome fell within the range.
3. After eight to twelve periods, count how often it did.

If the outcome almost always falls within your 80 percent range, your range is too wide and therefore tells you little. If it falls within only half the time, your range is too narrow and you are projecting more certainty than you have. The second is far more common.

Look at the direction too. If the outcome keeps falling below the range, you do not have a spread problem but a bias problem: your forecast is structurally too optimistic.

Confidence per deal, per signal, per forecast

The word confidence appears at three levels, and they are often confused.

• Per deal: the probability that this deal will be won. A number between 0 and 100 percent.
• Per signal or recommendation: how sure a system is that a finding is correct, for example that a customer is paying too little or that a deal will slip. How to judge that is covered in how do you know whether an AI recommendation is reliable.
• Per forecast: the range around the total.

A system can be very confident per deal and still give a wide overall range, because a few large deals determine the outcome. Conversely, a forecast built from many uncertain small deals can have a narrow overall range, because the errors cancel each other out.

What makes the range wider

• Concentration: a few large deals or customers that determine much of the revenue.
• Short history: fewer periods from which to measure your errors.
• Change: a new product, a new market, a price change. The past then says less.
• Poor data: incomplete fields, dates that are not kept up to date.
• Long forecast horizon: a forecast three quarters out is always less certain than one for the current quarter.

A good forecast does not become narrower through cleverer calculation, but through less uncertainty in the input: a cleaner pipeline, more recurring revenue in view, large deals better qualified.

How to explain it to the board

Three rules that work in practice.

1. Always give three figures: bottom, expected, top. And state at what confidence.
2. Say what drives the bottom. "If the two largest deals slip, we land at EUR 1,850,000." That is a sentence someone can act on.
3. Show how well your previous ranges held. One line, such as "over the last eight quarters the outcome fell within the range six times", builds more trust than any explanation of the model.

How AI models estimate a probability per deal, and why that is exactly when you want to see a range, is covered in AI revenue forecasting explained. For the full picture of forecasting, see what is revenue forecasting.

Frequently asked questions

Which confidence level should you choose, 80 or 95 percent?

For day-to-day management, 80 percent is the most useful: wide enough to be honest, narrow enough to say something. A 95 percent range is often so wide that little can be decided from it, but it is useful for cash planning and conversations with the bank.

What if I have not kept historical forecasts?

Start keeping them now. Until you have enough periods, you can estimate a rough margin by forecasting past periods retrospectively with the data you had at the time, as far as you still have it, or work with scenarios.

Is a narrow range always better?

Only if it holds. A narrow range that is often missed is worse than a wide one that holds, because it leads to decisions based on certainty that is not there.

Can a range be asymmetric?

Yes, and it often is. If the base is fairly certain and the uncertainty sits mainly in a few large deals, the top is further from the expectation than the bottom, or the other way round. A system that always shows the same percentage above and below is calculating too simply.