Case study · TrueLift
TrueLift is the forecasting and decision tool I built to hold my own work accountable. This is one realistic B2B decision the way the tool walks through it: the forecast and its range, three options, the rule that picked one, what happened, and the grade at days 7, 14, 21 and 28. The numbers are illustrative. The method is the one that runs in the product.
history 90% band options what happened targetillustrative data
A quarter's pipeline target, six weeks of history, and a 28-day forecast with a 90% band. Three ways to spend: hold, push paid up 30%, or cut paid 25%. Each branch carries its own odds of clearing the target: hold spend 52%, push paid +30% 63%, cut paid −25% 21%.
The rule written down before the money moved was worst case: pick the branch whose bad outcome the quarter can survive. Push paid has the best odds (63%) but its downside is a spent budget with nothing to show. Hold wins at 52%. Expected value or minimax regret would be other legitimate rules; what matters is that the rule is recorded first.
The 28 days arrive one at a time. The line of what actually happened is drawn over the frozen forecast. Nothing is re-fitted after the fact; the band is the one that was promised on day zero.
Checkpoints at days 7, 14, 21 and 28. Day 7: inside the band (-6% vs the chosen branch). Day 14: inside the band (-6% vs the chosen branch). Day 21: outside the band (-31% vs the chosen branch). Day 28: inside the band (-6% vs the chosen branch). The day-21 miss stays on the record. A forecast that is never wrong was never a forecast.
Scorecard
A 90% band should catch reality nine times in ten. When it doesn't, that goes here too. Each row is a real forecast made before the window opened and graded after it closed.
No publicly graded forecasts yet.This table fills from TrueLift's forecast ledger as entries resolve. It will never be filled with illustrative numbers.
TODO (Martin): tag resolved ledger entries “public” in TrueLift (Decision atlas → Forecast journal) once the engine's /public/scorecard endpoint is deployed. Only grades and ratios are published; dataset names, slices and absolute volumes stay private.The assessment starts with a forecast of your main metric, with a range.