Sales Forecasting Methods: What Each One Trusts

Every sales forecast is a claim about the future built from records typed in by the people whose performance the forecast describes. That is not a criticism of anybody's honesty. It is a structural property, and choosing a forecasting method is mostly choosing which part of that problem you are willing to live with.

The published methods divide into a small number of families. Each one takes a different input, each fails in a characteristic way, and the right choice depends more on what your data can support than on which method is currently fashionable.

The families, and what each one actually trusts

Judgement-basedTrusts the seller

• Rep commit: each seller states what will close
• Manager adjustment: leaders discount or add on top
• Requires: that people are calibrated and incentives do not distort
• Fails when: a bad quarter makes optimism cheaper than the alternative

Pipeline-basedTrusts the stage

• Stage-weighted: value times a probability attached to each stage
• Requires: stages with exit criteria the buyer produces
• Fails when: stages advance on seller activity, so the arithmetic multiplies effort by a coefficient

History-basedTrusts the base rate

• Historical run rate, seasonally adjusted
• Length-of-cycle: deals dated forward from their creation date
• Requires: enough closed deals for a stable rate, and a stable motion
• Fails when: the market, the product or the team changed

A fourth family sits underneath the others in practice: bottom-up capacity modelling, which forecasts from inputs rather than from deals. Number of sellers, productive weeks, meetings each can hold, historical conversion, average deal value. It is the only method that works before there is a pipeline to read, which makes it the right instrument for a new segment or a new team, and the wrong one for a mature motion where the pipeline itself carries better information.

Stage-weighted forecasting and the assumption it hides

Weighted pipeline is the most widely used method and the one most often applied without its precondition.

The arithmetic is simple: multiply each deal's value by a probability attached to its stage, and add them up. The precondition is that stage membership is an observable fact, because the probability is derived from history and history is only a base rate if the population is comparable.

When stages advance on buyer evidence, the coefficient means something real: deals that reached this observable state closed at this frequency. When stages advance on seller activity, the same arithmetic computes the frequency with which reps who did a thing went on to close, which mixes deal quality and rep behaviour into one number and hides both. The output looks identical in either case, which is what makes this failure durable. Pipeline stages that earn their place covers how to write criteria that survive the test.

There is a second and quieter assumption. Weighted forecasting is only valid across a large enough population for the probabilities to average out. Applied to one seller with eight open deals, a seventy percent coefficient does not mean seventy percent of the value arrives; it means the largest deal either lands or does not, and the weighted figure will be wrong by a wide margin in whichever direction that goes. Weighted numbers are for territories and above.

The category error in most forecast reviews

The commonest forecasting failure is not a method problem. It is that two people in the room are forecasting different things and neither notices.

A commit number is a promise. A weighted number is an expected value. A best-case number is an upper bound. These are three different quantities, they will never agree, and a review where one person quotes commit while another quotes weighted produces a disagreement that feels like a judgement dispute and is actually a units error.

The fix costs one line in the reporting template: state which quantity every figure is, every time. Teams that do this find that most of their forecasting arguments were arithmetic misunderstandings, and the remaining ones are real and worth having.

Judging accuracy, and the metric that stops the gaming

Forecast accuracy needs its own measurement or the process improves nothing, and the choice of measurement decides what behaviour it rewards.

Measuring only whether the total landed rewards a forecast set low enough to beat, which is why sandbagged forecasts survive so long: they are accurate in the direction nobody complains about. Measuring absolute error in both directions removes that, because a forecast that comes in forty percent over is as wrong as one that comes in forty percent under.

Forecast process checklist

• ✓Yes: Every figure is labelled as commit, weighted or best case
• ✓Yes: Stage probabilities are derived from your own closed deals, not from a template
• ✓Yes: Accuracy is measured as absolute error in both directions
• ✓Yes: Deals with close dates inside a window shorter than the median cycle are flagged
• ✓Yes: The forecast is compared against a bottom-up capacity model at least quarterly
• ✓Yes: Definition changes are dated and shown on the same chart as the metric
• ~Depends: One deal can move the whole number, and it is named in the review

The last item belongs there as a disclosure rather than as a failure. Concentration is normal in mid-market and enterprise motions, and a forecast whose outcome rests on one opportunity should be presented that way rather than as a smooth aggregate that happens to contain a coin flip.

Choosing a method for your situation

The decision is mostly determined by two facts about your business: how many deals close per period, and how stable the motion is.

High volume with a stable motion favours history-based methods, because the base rate is genuinely a base rate and individual deals cannot distort it. This is where run rate and cycle-length forecasting are at their best, and where weighted pipeline is a reasonable cross-check rather than the primary.

Low volume with large deals favours judgement plus explicit deal-by-deal review, because no statistical method has enough events to work with. The discipline here is making the judgement inspectable rather than arithmetical: what has the buyer actually done, what is still unknown, who has not been met. That is what opportunity-planning methodologies exist to structure, and it is why the acceptance test at the front of the pipeline matters so much in this motion: with twenty deals a quarter, one wrongly accepted opportunity distorts the whole forecast. What qualified has to mean when money depends on it is the version of that definition we hold ourselves to.

A new segment, a new team or a new product favours bottom-up capacity modelling, because the historical rates you would otherwise use describe a motion you are no longer running.

Most companies of any size need two of the three at once, and the useful practice is to run the primary method and one independent cross-check, then investigate the gap rather than averaging it. Averaging two forecasts produces a number with no owner and no diagnosis attached.

The gap itself is the most informative output of running two methods, and it is usually thrown away. A weighted pipeline number well above a capacity model means either the pipeline is carrying deals that should have been disqualified or the conversion assumptions in the model are stale, and reading ten deals settles which. A commit number well below the weighted one means sellers are seeing something the stage coefficients do not, and the reasons they give are the most valuable qualitative data a forecasting process generates. Neither of those questions survives an averaging step.

Review frequency matters as much as method. A forecast reviewed weekly against a quarterly target produces twelve chances to notice a gap while there is still time to act on it, and also twelve chances to overreact to noise. The practical compromise most teams settle on is a weekly read of movement and pipeline created, with the full forecast revisited monthly, and the number that goes outside the sales organisation changed only when something real has changed rather than every time it wobbles.

What forecasting cannot do

Two things are routinely asked of a forecast and cannot be delivered by any method.

It cannot create pipeline. A forecast that shows a gap eight weeks before period end is reporting a targeting decision made a full cycle earlier, and the honest response is usually to protect the next period rather than to rescue this one. Compressing outbound into the gap produces rushed targeting and opportunities that inflate coverage without closing. Pipeline coverage works through the timing of that in detail.

It cannot survive a definition change silently. Move the boundary at which an opportunity enters the pipeline and several forecast inputs shift at once, in different directions, with nobody selling any differently. Date the change and put it on the chart.

Where we differ from standard practice

Much of the advice around forecasting assumes an outbound motion run the common way, and since this page sits on our site it is worth naming where ours differs and what that costs.

The standard assumption is that pipeline gaps are closed by increasing touches: more messages per prospect, spread over more weeks, with later ones landing in the same thread. We run one message per campaign, with no bumps and no thread replies. A non-responding audience becomes a new campaign built on a different premise rather than a reminder of the last one. The reasoning is mechanical: a follow-up reaches the population that already saw the message and chose not to answer, which is the population most likely to complain, and the reputation cost lands on the sending domain across everything else it sends. our write-up on why we stopped using follow-ups has the full argument. The forecasting consequence is real and worth stating: pipeline built this way tends to move steadily rather than in spikes, which forecasts better and fills slower.

We are also paid on attended meetings that meet criteria agreed in writing before launch, with budget, timing and authority deliberately outside the definition. That is the same instinct as dating a forecast definition: settle the ruler before anyone reads the number. The full argument, including what it costs us, is in why we stopped using follow-ups.

The short version

Forecasting methods divide into judgement, pipeline-weighted, history-based and bottom-up capacity models. Weighted pipeline is the most used and needs stages whose criteria the buyer produces, or it multiplies seller effort by a coefficient and reports the result as revenue. It also needs a population large enough to average, so it belongs at territory level and above.

Label every figure as commit, weighted or best case, because most forecast arguments are units errors. Measure accuracy as absolute error in both directions so sandbagging is penalised. Run one primary method and one independent cross-check, and investigate the gap rather than averaging it. And accept that no method creates pipeline: a gap found late is a message about next period.

Where the gap is genuinely supply, that is the half we run, on criteria agreed before launch: see what a first campaign produces.