Getting started
Get your first forecast
Scope a decision, read the preview skeptically, close the loop.
Goal
One forecast, classification, or recommendation you can defend: the result, the evidence behind it, a recorded review — and an operational outcome that feeds the next round. Done means you acted on it or deliberately did not, with reasons either way.
Image: forecast output with evidence links and review actions visible beside the figures.
Steps
- Start from solid ground: a connected, syncing source and — ideally — a shaped ontology slice. Predictions read state; weak state makes weak predictions, and no review step can fix inputs that were never there.
- Describe the decision, not the method. "Which of our top 20 accounts are at renewal risk next quarter?" beats any instruction about models or algorithms. Scope three things explicitly: what is being predicted, over which time range, and for whom the answer matters.
- Read the proposal as a scope check before anything runs: the predicted target, the data and time range it draws on, and what the output will look like. A forecast aimed at the wrong quarter or the wrong segment is a planning failure, not a model failure — fix it here.
- Dry-run and read the preview like a skeptic:
- Do the top drivers make business sense, or does the result lean on a proxy you do not trust?
- Is anything important missing from the inputs — a segment, a season, a recent change the data does not know about yet?
- Would you bet a decision on the spread, or only on the ranking? Uncomfortable answers go back as corrections to the scope, not forward as hopes.
- Commit, then open the result and walk the evidence chain: each figure back through transformations and steps to source records. If you cannot trace a number to its origins, treat the forecast as unfinished regardless of how plausible it looks.
Image: evidence chain expanded from one forecast figure down to source records and the run that produced it.
- Record your review: approve what is right, correct what is wrong, reject what misleads. A correction here is the highest-value feedback in the system — it teaches extraction, mapping, and prediction at once.
- Close the loop. Note the decision you made and watch the outcome: renewed or churned, resolved or escalated. The operational result is what turns this forecast into a better next one.
When forecasts disappoint
- Confident and wrong: check the evidence chain for a stale source or a shifted input the model could not see. Correct the record, not just the conclusion.
- Vague or hedged everywhere: the scope was too broad or the inputs too thin. Narrow the question or strengthen the state, then rerun.
- Right number, unusable output: the proposal aimed at the wrong audience or format. Rescope who the answer serves.
What good looks like
- You can explain the forecast to a colleague from its evidence, not its score.
- Corrections flow back instead of dying in a meeting.
- The next forecast on the same question visibly accounts for your feedback.
- Outcomes get recorded, so accuracy is measured, not assumed.