Ask a CRO how accurate their forecast is and you’ll usually get a number and a wince. The typical quarterly miss sits somewhere between 10 and 20%, and most of that error comes from one source: the forecast is built on fields a human filled in on a Thursday afternoon. When we started building Forecast in 2022, we set one rule. The model may read what reps type, but it is never allowed to believe it without corroboration.
Signals first, opinions second
Forecast is trained on behavioural Signals rather than stated intent. For every open opportunity, we assemble a timeline of what actually happened: who replied and how quickly, which stakeholders joined calls, whether legal or procurement appeared, and whether the billing system saw a quote turn into an invoice.
In practice, a handful of Signals do most of the work. Across a training set of 2.3 million closed opportunities, these five explained more variance in outcome than every CRM field combined:
Buyer-side response latency over the last 14 days, relative to that deal’s own baseline
The number of distinct buyer-side participants on calls, weighted by seniority
Whether an economic buyer has actually spoken — not just been cc’d
Days since the last mutual next step was agreed on a call
Procurement or security-review activity from the buyer’s email domain
Notice what isn’t on the list: stage, amount and the rep’s own probability. Those still feed the model, but as features to be checked against behaviour, not as ground truth. When a rep’s call and the Signals disagree, Forecast shows both — and names the Signals driving the gap.
Calibration is the whole game
A model that says 70% should be right about seven times out of ten. That sounds obvious, and it is exactly where most forecasting tools quietly fail. We recalibrate every workspace’s model weekly against its own closed outcomes, using isotonic regression on top of a gradient-boosted ensemble. Until a team has at least 150 closed opportunities to calibrate against, we show ranges instead of false precision.
Across 212 customers in the first half of 2026, the result was a median absolute error of 3.7% on the quarterly commit, measured on day 30 of the quarter. The 90th-percentile customer landed at 8.1% — still comfortably ahead of the typical human call.
Calibration also changes how teams behave. When reps know a 40% deal genuinely means 40%, they stop sandbagging it to 10% or inflating it to 80% for the pipeline review. Over two quarters, the gap between rep-submitted and model probabilities narrowed by more than half in teams that shared the calibrated number openly.
“For the first time, the number I take to the board is the number I actually believe.”
— Hanna Lindqvist, Chief Revenue Officer, Parcelwise
What we still get wrong
Forecast struggles in two places: brand-new segments, where there’s no history to calibrate against, and very large deals, where one champion changing jobs can swing a quarter. For both, we now show a wider confidence band and highlight the handful of deals that dominate the variance, so your team spends its inspection time where it actually moves the number.
We’ll publish the full methodology, including our evaluation harness, later this year. If you want to poke holes in it before then, our inbox is open.
