Sales Forecast Calibration

A team labels 100 deals as having an 80% chance to close this quarter, and if only 40 close across several comparable quarters, the label is poorly calibrated. The team should inspect stage rules, dates and customer mix before changing every probability mechanically.

Define the forecast target first, since closed bookings, recognised revenue, invoiced sales and cash are not interchangeable, and a probability of closing by quarter-end differs from a probability of ever closing. Save the prediction as it was made, because if a manager revises a deal from 20% to 90% the day before signing, using the final version to judge earlier accuracy hides the original decision risk.

Salesforce's forecasting guidance discusses comparing forecasts with actual sales over time and examining why forecasts miss, though its example performance figures are vendor experience, not a benchmark every team must meet. CFA Institute's forecasting guidance discusses drivers, assumptions and scenario analysis, so calibration should revisit the inputs that fit a business model rather than impose an arbitrary target for every product.

Group similar predictions, so that for deals labelled 70-80% likely you compare observed wins after the stated horizon, remembering that a handful of deals is too noisy to draw a strong conclusion. An illustrative calibration gap for one probability group is observed win share minus average predicted probability: if deals averaged 80% predicted and 50% won by the stated date, the gap is negative 30 percentage points, which does not explain why.

Check value as well as count, because a model may predict wins accurately by count but miss revenue if the largest opportunity slips. Inspect timing errors, since a deal that closes next quarter is not a win for the current-quarter forecast, and track delayed wins separately from permanently lost deals.

Review stage definitions, because reps may use 'proposal sent' differently, and account for changes in the pipeline such as a new market, price, product or sales cycle, backtesting on recent, relevant cohorts and stating when there are too few observations. Track bias, as forecasts may consistently overshoot because close dates slide or discount approvals take longer than expected, and measure dispersion, since two forecasts with the same average error can differ if one has occasional very large misses.

Use human review carefully: a manager can override a model with new buyer evidence, but the reason and time should be recorded so an unsupported optimistic override is visible in the later review, and honest uncertainty should not be punished, since incentives that reward certainty or pipeline size invite gaming. Test changes prospectively, because improving a backtest alone is not proof the new approach works, and segment the test where there are enough deals, as large enterprise deals may close differently from smaller renewals.

Avoid hindsight bias by keeping both the forecast made at the time and the outcome, so no one can rewrite expectations after a deal slips; for an owner, calibration means a forecast label has a track record behind it.