The Real Ceiling on Sales Forecast Accuracy
Sales forecast accuracy has a real ceiling, and it sits higher than most revenue teams believe. In our data at ORM, forecasts on new and expansion revenue land around 90% accurate, and a properly trained model holds that from day one of the quarter to day ninety. The teams that miss are not missing because the ceiling is low. They are missing because they buy their accuracy with manual effort that decays the moment the market moves.I have built forecast models for B2B SaaS companies, and the question I get most is some version of "what accuracy is even possible?" People ask because they have been burned. They tightened their revenue forecasting process, hit their number one quarter, missed by fifteen points the next, and concluded that forecasting is closer to astrology than math. It is not. There is a knowable ceiling, there is a cost to reaching it, and there is a path. This post is all three.
What "Accurate" Actually Means
First, a definition, because most arguments about accuracy are really arguments about scope.
When I say roughly 90% accuracy is realistic, I mean on new and expansion revenue, forecast as its own motion. Renewal is a different animal with different mechanics, and blending it into one number is how teams fool themselves. If you want a defensible forecast accuracy figure, you separate the motions and score each on its own.
Second, accuracy is not a single end-of-quarter grade. A forecast that becomes accurate in the last week of the quarter is worthless, because by then the quarter has already happened. The accuracy that matters is the accuracy you have on day one, and whether it holds as the weeks pass. Those are two different bars, and most teams only clear the first one by accident.
What 90% Actually Costs
Here is the part nobody tells you. A team can reach 90% accuracy on new and expansion with a manual process. I have seen it. But it is expensive in a specific way: it takes enormous effort to produce, it is static the day it ships, and it does not react as conditions change.
That last point is the mechanism behind most misses. A forecast fails when something in the business or the market changes and the forecast is still built on last quarter's assumptions. A new competitor enters and average deal size drops. Interest rates rise, buyers slow down, and win rates fall. Uncertainty stretches sales cycles from qualified to closed. You reshuffle territories and reps get distracted, so execution slips even while pipeline looks healthy. A manual forecast cannot feel any of that in time. Someone has to notice, quantify it, and rebuild the model by hand, and by the time they do, the quarter has moved again.
This is why I distrust a forecast that leans on pipeline coverage as its headline. Coverage is an input, not an answer. In our data, the average deal size in pipeline is $80,000 while the average closed-won deal is $40,000. A team holding 4x coverage on those inflated values is not covered at all. It is looking at a number that makes executives feel informed while masking the actual risk.
The Forecast Accuracy Ladder
I use a four-rung model I call the Forecast Accuracy Ladder. Each rung is a prerequisite for the one above it. Skip a rung and the accuracy above it is luck, not a system.
Rung one: consistent data. Not clean data. Consistent data. If your inputs are wrong in the same way every time, a model can learn the bias and correct for it. This rung is where most teams think they are stuck, and they are almost never actually stuck here. Rung two: a model trained on your history. At ORM, a fully trained model built on a company's historical sales performance takes 4 to 6 weeks. That is the real cost of entry, and it is a one-time cost, not a recurring tax. The model learns how your business closes: which deals group together, how long each group takes, what your win rates actually are once you strip out the optimism. Rung three: deal grouping and aging curves. We group every opportunity with a machine learning model, and for each group we predict a curve for how long it takes to close. Those curves run from 1 to 80 weeks, with most of the expectation landing before week 12 and very few groups closing past 52 weeks. This is what lets a forecast know that a deal is aging out before a human notices. We treat a change in stage, close date, or amount as meaningful activity. The absence of all three is a warning. Rung four: continuous re-forecasting. The model updates as the quarter progresses so you always hold the most accurate read. This is the rung that turns a static 90% into a dynamic 95%. It is also the rung that catches deal slippage early, because the single clearest slippage signal is a rep moving a close date, and a live model reprices that deal the instant it happens.| Manual quarterly forecast | Trained dynamic model | |
|---|---|---|
| Accuracy on new + expansion | ~90% | ~95% |
| Holds day 1 to day 90 | No, drifts as conditions shift | Yes, re-forecasts continuously |
| Reacts to market change | Slow, requires a manual rebuild | Automatic |
| Effort per cycle | High, every quarter | Low, largely automated |
| Setup cost | Ongoing forever | 4 to 6 weeks, once |
A Worked Example
Consider Halden Systems, a mid-market SaaS company with a healthy-looking pipeline. Numbers here are illustrative, not a benchmark.
Halden starts a quarter with a $9M new-and-expansion pipeline against a $2.4M target. That is 3.75x coverage, comfortably inside the standard band. The VP of Sales calls it a strong quarter on day one.
Then you run the composition instead of the ratio. Half the pipeline value sits in five deals owned by two reps. A third of it has not had a stage, amount, or close-date change in over a quarter, so the aging curves flag it as unlikely to close in-period. And the pipeline's average deal size runs double what Halden actually closes at. Reprice the pipeline on real close values, strip the stale deals, and concentrate-deal risk, and the credible in-quarter number is closer to $2.0M. The 3.75x coverage was real. The forecast it implied was fiction.
That gap does not show up in a coverage ratio. It shows up the moment a trained model decomposes the quarter into what will actually close, what still has to be created and closed in-period, and what might get pulled forward at a discount. Halden did not have a data problem. It had a model that read the pipeline it could see and ignored the motion it could not.
The Contrarian Part
Here is the position I will defend, and it is the one that gets the most pushback: you do not need clean data to forecast accurately, and waiting for clean data is the most expensive form of procrastination in revenue operations.
Every team I meet believes its data is uniquely bad, and that this is the reason they cannot run the business the way they want. It is not true. Everyone has messy data. It does not matter. As long as the mess is consistent, a model can predict through it. The teams stuck at 70% accuracy are not stuck because of dirty CRM fields. They are stuck because their forecast is a static spreadsheet that cannot react, or because they are grading themselves on end-of-quarter accuracy that arrives too late to act on. Chasing perfect data is a way of avoiding the harder work, which is building a model that re-forecasts itself and decomposes the quarter before it starts.
The same logic applies to how you weight your inputs. AI forecasting accuracy does not come from feeding an LLM your pipeline and asking it to guess. It comes from machine learning and optimization trained on your specific history, with every number pointing back to a traceable source. When you can hold roughly 90% on new and expansion revenue from day one, and push toward 95% as the quarter runs, the value is not the final score. The value is knowing the likely shape of the quarter early enough to change it.
ORM trains that model on your history in about four to six weeks, then keeps it accurate on its own, so the forecast you trust on day one is still the forecast you trust on day ninety.
Frequently Asked Questions
What is a realistic sales forecast accuracy for B2B SaaS?
On new and expansion revenue, around 90% accuracy is realistic and achievable. In our data at ORM, a dynamic model trained on a company's history holds close to 95% without manual adjustment. Renewal is a separate motion and should be forecast on its own, not blended into the same number.
How long does it take to train a forecasting model?
At ORM, a fully trained model built on a company's historical sales performance takes 4 to 6 weeks. The model learns how that specific business closes: deal groupings, aging curves, and win-rate patterns. After that it re-forecasts continuously as the quarter progresses, so accuracy holds from day one to day ninety.
Why do sales forecasts miss even with strong pipeline coverage?
Pipeline coverage is an input, not a forecast. A team can hold 4x coverage and still miss if the pipeline is stale, concentrated in a few large deals, or priced above what deals actually close for. In our data, pipeline deals average $80,000 while closed-won deals average $40,000, so coverage measured on inflated values masks the real risk.
Do you need clean data to forecast accurately?
No. Garbage in does not have to mean garbage out. As long as your data is consistently wrong in the same way, a model can learn the pattern and correct for it. Every revenue team believes its data is uniquely bad, and it almost never is the thing blocking an accurate forecast.
What is the earliest signal a deal will slip?
The clearest slippage signal is a sales rep changing the close date. Once a deal moves from one quarter to the next it is less likely to close, even when it sits in commit. The earliest signal is the absence of activity: no stage change, no amount change, no notes, and no buyer response.
Should renewal revenue be forecast the same way as new business?
No. New and expansion revenue behave differently from renewal, so blending them into one accuracy figure hides where you are actually strong or weak. The roughly 90% accuracy benchmark applies to new and expansion. Forecast renewal on its own retention mechanics and report the two separately.
See how ORM turns these insights into action
ORM builds custom revenue forecast models for B2B SaaS companies. Not dashboards. Prescriptive analytics that tell you what to do next.
Schedule a Demo