Sales Forecasting Methods: Run Two, and Let Them Disagree

Somebody in your company commits to a number every quarter, and almost nobody in the room believes it to the nearest ten per cent. That is not cynicism about your team. It is the normal condition of a discipline that asks one model to predict human behaviour and then gives it no way to check itself.

What follows is the standard menu of forecasting approaches, the specific way each one breaks, and a rule that costs nothing to adopt: run two of them, chosen so they fail differently, and treat the gap between their answers as the output.

The Miss Is Neither Rare Nor Small

Xactly's 2024 benchmark survey of 405 finance, sales and revenue operations professionals put numbers on how far off the typical commit sits. CFO.com reported that 81% of those leaders said their forecasts are usually at least 5% out, and 43% said the usual error is 10% or worse. The figures are two years old now, and we are using them because they come with a stated sample and independent coverage, which is more than most numbers circulating on this subject can claim.

The same coverage carries a detail that explains more than the headline does. Ninety-five per cent of those finance and revenue operations teams said they were confident planning on the forecasts they already have. Read those two findings side by side and you get the real problem: the error is routine, it is known, and it is not treated as information.

The causes Xactly names are unglamorous. Systems that cannot reach the data, leaders arriving unprepared at the forecast call, reps who were never told which deals belong in the number, and pipeline hygiene nobody owns. Not one of those is solved by a model.

Six Approaches, and the Question Each One Answers

Michael Duncan, who works in go-to-market strategy at Gong, sets out six approaches in the company's guide to forecasting methods. The taxonomy is a good one, so here it is in our words, with the question each approach is genuinely answering.

1. Opportunity stage. Multiply each open deal by its stage's historic conversion and add it up. Asks: given where everything sits today, what should land?
2. Intuitive. The judgement of the people closest to the deals. Asks: what do the reps and their manager actually believe?
3. Historical. Last year, plus growth, plus seasonality. Asks: if the pattern repeats, where do we end up?
4. Multivariable. Many inputs at once, from call content to market conditions. Asks: which signals have predicted closure in our own data?
5. Regression. Tie volume to a driver such as spend or headcount. Asks: how far does one input move the outcome?
6. Cycle length. Judge each deal by its age against the usual duration. Asks: can this one finish inside the period at all?

Treat that as a menu, not a decision. Nothing about the list says you may order once.

Every Method Has a Blind Spot It Cannot See

Each approach fails in a way that is specific and predictable. Stage-based numbers collapse when stage definitions are soft, because the probabilities then measure how a rep was feeling on Tuesday. Intuitive numbers inherit whatever mood the floor is in. Historical numbers hold until the market moves, at which point they hold confidently and wrongly. Cycle-length numbers break on the unusual deal, which is frequently the largest one.

The structural point is that a method cannot detect its own failure mode. Nothing inside a stage-weighted calculation can tell you the stages have drifted. You need a second instrument, built on a different assumption, and then the disagreement between them does the detecting for you.

Before any of this, fix the definitions. Write the evidence required to enter and leave each stage in a single line, something testable such as a named economic buyer or a dated plan the buyer has agreed. Without that, you are not forecasting, you are aggregating opinion, and our walkthrough of pipeline hygiene practices is the first pass to run.

Why Two Numbers Beat One Better Number

This is not a sales idea, and that is the strongest argument for it. Combining forecasts is one of the oldest results in the forecasting literature. In their review of the field, Xiaoqian Wang and co-authors including Rob Hyndman describe combination as a way of integrating information from different sources while "mitigating the risk of identifying a single 'best' forecast". The risk being mitigated is precisely the one a sales team runs every quarter: picking one model, trusting it, and finding out at the end of the period which assumption it was quietly making.

Pair methods whose weaknesses do not overlap. Stage-weighted against cycle length is a good pair, because one is blind to timing and the other is built on it. Historical against intuitive is another, because one cannot see a changing market and the other sees little else. Record both numbers before the quarter opens, in writing, and do not recalculate either one in week nine.

Then use the gap. If one method says the quarter is comfortably covered and the other says it is short, the deals that explain the difference are the agenda for the weekly call. That is a far more productive meeting than a negotiation over whose optimism is more credible, and it pairs naturally with getting the review window right, which our piece on the week your pipeline review is actually measuring takes apart.

Accuracy Is the Wrong Target. Explainability Is the Right One.

Here is where we part company with most of the material on this subject, including the vendor guides that list accuracy percentages for each method as though they were specifications. The position: a forecast's job is not to be right, it is to be wrong in a way you can explain before the quarter ends. A team that misses by a tenth and can name the three deals and the wrong assumption is in far better shape than a team that lands on the number by accident and learns nothing it can reuse.

That reframing changes what you optimise. It makes the gap between two methods the deliverable rather than an embarrassment to be reconciled before the board sees it. It makes rep judgement a recorded input sitting beside the model rather than something to be overridden, which is also why resistance tends to fall away once sellers see their call written down rather than deleted. And it kills the search for a single method with a published accuracy figure, because that figure was measured on somebody else's pipeline, in somebody else's market, by somebody who sells forecasting software.

Very few deals a quarter leaves this with less to say. Run eight opportunities and any method is dominated by whether one of them slips, two methods will mostly agree on the obvious and disagree on noise, and the honest answer is a deal-by-deal conversation rather than a model at all. The threshold is roughly whether you have enough closed history for a stage conversion rate to mean anything, and if you do not, saying so out loud is the right move, as we argued in not enough data yet being the most honest thing a pipeline can say.

Back-Test Before You Argue

You can settle this without a debate, using data you already have. Take last quarter's opening pipeline, apply two methods to it, and compare each against what actually closed. One of them will have beaten whatever you currently report. That result is your case to leadership, and it took an afternoon.

Four habits make it stick. Score both methods at every quarter end and keep the error by method, not just the error overall. Retire a challenger only after three quarters, because one is noise. Keep the coverage assumption visible next to the forecast, since a number built on an unexamined conversion rate is the same mistake in a different column, which is the argument in our piece on deriving coverage from your own win rate. And expect the first quarter to feel worse, because two numbers that disagree in public are uncomfortable in a way one confident number never is.

No method removes uncertainty, and nobody can promise you an accuracy figure. What a second method buys is an early warning and a vocabulary for the miss. On a function that runs on commitments made in advance, that is worth more than another dashboard.