Most forecasts are produced by negotiation up a management chain, then defended as arithmetic. A defensible forecast is one where every delta from last week is attributed to a specific deal and a specific event. Here is how to build one.
Ask five people in the same company what the quarter will land at and you will get five numbers, each defensible in its own frame. The rep's number is a commitment. The manager's number is the rep's number with a haircut applied from memory. The VP's number is the manager's number adjusted for what the VP believes about the manager. The CRO's number is what can be said out loud. None of these are predictions in any technical sense.
This is not dishonesty. It is what happens when a forecast is produced by a chain of humans each correcting for the bias of the one below, without anyone writing the corrections down.
What makes a forecast defensible
A forecast is defensible when three questions have answers that survive being checked:
- What is the number, and what is the uncertainty around it?
- Why did it change since last week — which deals moved, and what happened to them?
- What would have to be true for it to be wrong, and how would we know early?
Most forecasting processes answer the first question and treat the other two as a conversation. The second question is the one that actually matters, because a number you cannot decompose is a number you cannot act on.
Close dates are commitments, not predictions
A rep-entered close date encodes intention, quota pressure, and what the rep believes their manager wants to hear. It is a useful signal — it tells you what the person closest to the deal is willing to commit to — but it is not an estimate of when the deal will close, and treating it as one is the single largest source of systematic error in CRM forecasting.
The correction is empirical and unglamorous: for each segment, compute the historical distribution of the gap between first-committed close date and actual close date. If deals in a segment slip a median of two weeks and the distribution has a long right tail, that is a prior you can apply to every open deal in that segment without asking anyone to change behaviour.
The forecast does not need reps to be more accurate. It needs the system to know how inaccurate they are, and in which direction.
The premise of every calibrated model
Variance attribution is the whole product
The most valuable thing a forecasting system can produce is not the number. It is the diff.
Last week the quarter was at a certain figure. This week it is lower. A defensible system tells you: four deals slipped out of the quarter (named), two of them because procurement engaged later than the segment norm; one deal was upgraded in confidence because the economic buyer joined a call; one new deal entered commit. The net of those movements is the delta. Nothing is unexplained.
Once that exists, forecast review stops being a negotiation about the number and becomes a conversation about the deals — which is the conversation that actually changes outcomes.
Categories should be defined by evidence, not by feel
Commit, Best Case and Pipeline are useful categories that almost nobody defines. Write the definitions down as checkable conditions:
- Commit — economic buyer identified and engaged, mutual close plan in writing, pricing agreed in principle, no open blocking dependency.
- Best Case — active and progressing, but at least one of the above is unresolved.
- Pipeline — everything else, including deals that look good and have not yet been tested.
The definitions matter less than the fact that they are written and checked. A system that flags 'this deal is in Commit but has no identified economic buyer' does more for forecast accuracy than any model, because it corrects the input rather than modelling around the error.
Judge the system on calibration, not accuracy
A forecast that hits the number one quarter and misses by twenty percent the next is worse than one that is consistently three percent low, because the consistent one can be corrected and the erratic one cannot.
The measurement to track is calibration: of everything the system called at seventy percent confidence, did roughly seventy percent close? Plot it every quarter. A well-calibrated system that is often uncertain is far more useful than a confident one that is occasionally spectacular.
What to do with a number you now trust
The point of a defensible forecast is not the reporting. It is that a number with attached reasoning can be acted on early.
If the model says the quarter is short and the shortfall is concentrated in one segment where deals are slipping past a procurement step, that is a specific, addressable problem in week three — not a surprise in week twelve. The forecast stops being a scoreboard and becomes an instrument.
The uncomfortable part
A defensible forecast will sometimes tell you the number is lower than the one already communicated. Every organisation that builds one has to decide, once, whether it wants an instrument or a scoreboard. Almost all the value is on the instrument side, and almost all the discomfort arrives in the first quarter.