Sales forecasting methods that actually work
The method you should use is decided by two things: how many open deals you have, and whether you keep a record of past forecasts. With many small comparable deals, historical rates work. With few lumpy ones, only a range does. Without a record, every method is guessing with extra steps, because nothing tells you whether it is working.
Below, six methods in the order most teams meet them, what each is genuinely good for, and where it breaks.
1. Stage-based (weighted pipeline)
Each stage carries a probability; multiply by value and sum. Universal because every CRM does it by default.
Good for: high-volume, repeatable sales where deals are similar in size and the stage genuinely reflects buyer progress. Breaks when: you have fewer than about thirty deals, or one deal is much larger than the rest. It is an average, and an average of eleven unequal things describes nothing.
2. Rep judgement (commit, best case, pipeline)
Each rep sorts their deals into categories and the manager rolls it up. Usually combined with a forecast call.
Good for: complex sales where the person on the deal knows things no field captures. Breaks when: nobody scores the judgement afterwards. Unscored judgement drifts optimistic, because being wrong has no cost and sounding confident has a reward.
Judgement is the best input available and the worst one to leave unmeasured. A rep's 70% is valuable the moment you know whether their 70% closes seven times in ten.
3. Historical rate
Take what you closed in the equivalent period last year, adjust for growth and seasonality. Ignores the current pipeline entirely, which is a feature as much as a flaw.
Good for: a sanity check against pipeline-based forecasts, and for businesses with stable, repeating demand. Breaks when: anything structural changed — a new product, a new market, a team that doubled. It assumes next quarter resembles last year.
4. Pipeline velocity
Number of opportunities, average deal value, win rate and sales cycle length combined into a revenue-per-day figure. Useful as a diagnostic rather than a forecast.
Good for: spotting which of four levers is degrading, and comparing teams. Breaks when: used to predict a specific quarter. It is an average of averages, so it is even less sensitive to a single large deal than weighted pipeline is.
5. Regression and machine learning
Fit a model to historical deal attributes and let it score open opportunities.
Good for: organisations with thousands of closed deals and consistent data capture. Breaks when: you have hundreds rather than thousands, or your CRM fields are filled in inconsistently — which is nearly everyone. A model trained on sparse, optimistic data produces confident, optimistic predictions, and the confidence is the dangerous part.
6. Simulation (a range rather than a number)
Treat each deal as a weighted coin, play the quarter out thousands of times, and report the distribution: where you are likely to land, the odds against target, and which deals carry the uncertainty.
Good for: exactly the case the other methods handle worst — few deals, uneven sizes, where the question is "what are the odds" rather than "what is the number". Breaks when: your input probabilities are rubbish. Simulation propagates your judgement honestly; it does not improve it.
How to choose
- More than ~30 comparable deals, stable business: stage-based is fine. Sanity-check against historical rate.
- Fewer than ~20 deals, or one deal over a fifth of target: use a range. Averages cannot describe your quarter and will report a number you have little chance of hitting.
- Any size, no record of past forecasts: start the record this month, whatever else you do. It is twelve rows a year and it is what makes every other method improvable.
The methods are not really in competition. Most teams should run two: one that uses the pipeline and one that ignores it, and treat a large gap between them as the signal it is.