Sales Forecasting Methods: The Six That Matter, and When to Use Each

A sales forecasting method is the logic you use to turn what you know into what you expect, the rule that converts pipeline, history, and deal attributes into a number you are willing to commit to. Every method makes the same promise. They differ entirely in what they read to keep it.

I have built revenue and forecast models for B2B SaaS companies for two decades, and I will open with the claim this whole guide defends: the method you pick matters far less than whether it fits the data you actually have. The industry talks about forecasting methods as a ladder, with spreadsheets at the bottom and AI at the top, and treats moving up the ladder as moving toward accuracy. That is wrong, and it is an expensive kind of wrong. Only 7% of companies achieve 90%+ forecast accuracy (Gartner), and plenty of the teams missing that bar are running sophisticated tools on data those tools cannot trust. Sophistication is not accuracy. Fit is accuracy.

So this is not a ranking. It is a tour of the six methods that matter, what each one reads, where each one fits, and a test for choosing among them that has nothing to do with which sounds most advanced.

The six methods, and what each one actually reads

The clearest way to separate the methods is not by how complex they are but by what input they take. Seen that way, the right choice for a given team stops being a matter of taste.

Method	What it reads	Fits best when	Main weakness
Historical	Your own past revenue, trended forward	Revenue is stable and seasonal, motion is steady	Blind to anything new in the pipeline
Pipeline stage-weighted	Open deals times a fixed stage probability	You need a fast, CRM-native number	Fixed percentages rarely match real conversion
Opportunity scoring	Per-deal attributes scored for likelihood	Reps disagree with the stage math	Only as good as the inputs reps log
Regression	Historical variables fit to an equation	You have clean history and want to know which drivers matter	Assumes the future behaves like the past
AI / machine learning	Every signal at once, patterns learned	Years of clean data and real deal volume	Learns your bad data if the data is bad
Prescriptive	The forecast plus the levers that move it	You need a decision, not just an estimate	Requires the model to reach into the motion

Read the table as a sequence of inputs, not a quality order. Historical reads the rear-view mirror. Stage-weighted and scoring read the pipeline in front of you. Regression and AI read history and attributes together. Prescriptive reads all of it and then reads one thing more, the set of actions available to you, which separates a forecast you watch from a forecast you act on. Each method is the right answer for some team. None is the right answer for every team.

Historical

The oldest method and the most underrated. You take your own past, adjust for trend and seasonality, and project it forward. It works astonishingly well when revenue is stable and the motion has not changed, because the best predictor of a steady business is a steady business. It fails the moment something new enters the pipeline that the past never saw, a new segment, a new competitor, a new product line. Use it as your baseline, the starting point any guide on how to create a sales forecast builds from. If a fancier method disagrees with a clean historical trend, the fancier method owes you an explanation.

Pipeline stage-weighted

The default in nearly every CRM, and the most overrated of the six. You assign each stage a fixed probability, multiply every open deal by the probability for its stage, and sum. A deal at the 60 percent stage worth 100K contributes 60K. The appeal is that it is instant and native. The problem is that those fixed percentages are almost always fiction. The deals sitting in your "60 percent" stage do not convert at 60 percent, they convert at whatever your actual stage history says, and that number drifts by segment and over time. Stage-weighted forecasting is fine as a rough coverage read. Trust it as a commit number and it will flatter you right up until the quarter closes short. If you lean on it, at least calibrate the percentages against real conversion, which is a sales process optimization exercise as much as a forecasting one.

Opportunity scoring

Instead of trusting the stage, score the deal. Each opportunity gets a likelihood built from its own attributes, engagement, champion strength, deal age, fit, budget confirmation. It is a real improvement on stage math when reps know the percentages lie, because it judges deals on what is true about them rather than which column they sit in. Its weakness is the same as its strength: it is only as good as the inputs, and inputs come from reps logging fields under deadline. Good scoring with disciplined data entry beats stage-weighting handily. Sloppy scoring is stage-weighting with extra steps.

Regression

Now you leave the pipeline and mine the history. Regression fits past outcomes to variables, lead source, deal size, segment, season, and hands you an equation that says which drivers actually move revenue and by how much. Its real value is explanatory: it tells you that enterprise deals from one channel close at twice the rate of another, which changes where you invest. Its limitation is structural. Regression assumes the future behaves like the past, so the more your market is shifting, the more carefully you have to hold its output. It needs clean history to mean anything, which is exactly why it belongs downstream of the data discipline in a forecast accuracy guide, not upstream of it.

AI and machine learning

This is the method everyone wants and most teams are not ready for. Machine learning reads every signal at once, pipeline, history, attributes, engagement, and learns patterns no human would hand-code. The ceiling on accuracy is genuinely higher here. So is the floor on data quality. A model fed clean, consistent, multi-year data finds real patterns. The same model fed inconsistent stages and half-empty fields learns your data-entry habits and reports them back with total confidence, which is worse than a simple model that wears its limits on its face. AI raises the ceiling. It does nothing for the floor. Earn the floor first.

Prescriptive

The other five methods stop at an estimate. Prescriptive forecasting does not. It produces the number, then names what to do about it, which deals to work, where coverage is short, which lever closes the gap between the forecast and the target. It reads everything a machine-learning model reads and adds the one input the others ignore: the actions available to you this quarter. The catch is that prescriptive only works when the model is wired into the motion deeply enough to recommend a move and watch it land. Done shallow, it is predictive with a to-do list bolted on. Done right, it is the only method that ends at a decision. For the full definition, the prescriptive analytics glossary entry draws the line precisely.

The Forecast Fit Test

Here is the framework I use when a revenue leader asks which method they should run. I call it the Forecast Fit Test, and it deliberately refuses to ask which method is most advanced. It asks three questions in order, and the order is the point, because failing an earlier question disqualifies every method that depends on passing it.

1. How clean and deep is your data? This gates everything above the simplest methods. If your CRM history is inconsistent, your stages mean different things to different reps, or you have less than a couple of years of trustworthy data, regression and machine learning are off the table no matter how much you want them. They will learn your mess. A team in this position is better served by a disciplined historical model and a calibrated stage read than by an AI system it cannot honestly feed. Clean the data first. Then climb. 2. What is your deal volume and shape? High-volume, lower-value motions give patterns enough repetitions to be real, which is where scoring and machine learning earn their keep. Low-volume, high-value enterprise motions, a handful of large deals a quarter, do not, because every deal is an outlier and the model is fitting noise. Those motions live closer to opportunity scoring and human judgment than to statistics, and pretending otherwise is how you get a confident forecast built on six data points. Match the method to how many deals you actually close, a constraint the sales forecasting complete guide lays out as the pillar around this whole cluster. 3. What decision does the forecast have to drive? A forecast that only has to inform the board can stop at a good predictive number. A forecast that has to tell a sales leader where to spend the next nine weeks has to reach further, into the prescriptive layer, or it is not doing the job. Decide what the forecast is for before you decide how to build it. The most accurate estimate in the world is worthless if it arrives without the decision it was supposed to trigger.

Run those three in order and the right method usually names itself. Notice what the test never asks. It never asks which method is newest, and it never lets ambition skip a step. That omission is the entire argument of this guide.

A worked example: Marlowe Grid picks a method

Numbers here are illustrative, not a benchmark. They exist to show the test deciding.

Marlowe Grid is a B2B SaaS company that just crossed into the mid-market and wants to "upgrade to AI forecasting" after a board member asked why they were not using it. The instinct is to climb straight to the top of the ladder. The Forecast Fit Test says stop and check the rungs.

Question one, data. Marlowe Grid has eighteen months of CRM history, and a migration nine months ago renamed three pipeline stages, so the data before and after the migration do not mean the same thing. That fails the cleanliness gate outright. Machine learning fed this history would learn the migration, not the market. Regression would fit an equation to a definition change. Both are off the table until the data is reconciled, which is roughly a two-quarter project, not a tool purchase.

Question two, volume. Marlowe Grid closes a moderate number of mid-market deals a quarter, enough for stage-level conversion to be meaningful when calibrated, but not enough for a learned model to separate signal from noise even if the history were clean. The volume points to a calibrated stage-weighted or opportunity-scoring approach, not statistics.

Question three, decision. The forecast has to do two jobs: give the board a credible number and tell the VP of Sales which deals to defend in a quarter where median B2B win rates have fallen to 19% (First Page Sage, 2025) and sales cycles have lengthened 22% since 2022 (Digital Bloom, 2025). The board job a stage model can do once calibrated. The "which deals to defend" job is prescriptive, and no amount of historical math reaches it. Pinning that decision to a live number is where the sales KPIs that steer a quarter actually earn their place.

So the right answer for Marlowe Grid is not the AI system the board asked for. It is a calibrated stage-weighted model today, with opportunity scoring layered on as rep-logged data improves, and a clear plan to reconcile the history so that a learned model becomes honest in a year. The same company that walked in asking for AI walks out with a sequence, because the test graded its data before its ambition. Jump straight to machine learning and Marlowe Grid would have bought a confident forecast trained on a stage migration, which is exactly the kind of thing that lands a team inside the 87% of enterprises that missed revenue targets in 2025 (Clari Labs, 2026).

Build the staircase, don't jump to the top

If you take one thing from this, make it the sequence. Forecasting methods are not a ladder you climb for status. They are a staircase you climb as your data earns each step. The teams that miss are rarely the ones running a method that is too simple. They are the ones running a method too advanced for the data underneath it, getting a number that looks authoritative and is quietly wrong, the most dangerous forecast there is, because nobody questions it until the quarter is already lost. The honest move is to run the most advanced method your data can actually support, no further, and spend the saved energy making the data clean enough to earn the next step. ORM's work sits at the top of that staircase, in the prescriptive layer where 95%+ forecast accuracy comes not from a fancier model but from feeding the right method data it can trust.