Predictive BI
The dashboard that sees a little ahead
Predictive BI marries statistical models to the dashboard: forecasts with ranges, lead-time steers, and claims you can justify. Prediction becomes a steering dial, not a crystal ball.
What you will learn
- Frame prediction questions that a model can honestly answer
- Read and present forecasts as ranges, not points
- Keep predictive BI on the steering dial, off the pedestal
What is this?
Predictive BI adds a forecast or a score to the report stack: next-month revenue with a range, likely-churn accounts, pipeline reconversion estimates. The discipline is treating the prediction as an instrument — a range with assumptions, refreshed and validated — that steers the leading dials rather than being worshipped as destiny. It converts 'what happened' reporting into 'what's next, and what moves it'.
Why do you need it?
- Leads can act on a range long before an outcome lands.
- Predictive scores target care: churn-risk lists beat average churn rates.
- Forecasting is a natural share of every senior BI remit.
- It builds the bridge between BI and the analytics science team.
Mental model
Weather forecasts on the news
Predictive BI
Weathercasters say '70% chance of rain, 20mm possible' — a range and a probability, and we plan umbrellas anyway. Predictive BI should talk the same way: point plus interval plus confidence, meant for steering, never for a guarantee. Nobody sues the BBC for a wrong forecast; we do get mad when the dashboard overclaims certainty.
How it works
- Pick the forecastable: a metric with trend + seasonality + history.
- Model choice: baseline, exponential smoothing, regression, or a time-series model — sized to the question.
- Interval production: the model emits a range with its confidence (not a hard point).
- Refresh + validate: re-run on new data; check error metrics against the naive baseline.
- Steer: surfacing the range as a leading dial with assumptions labeled.
From history to steering forecast.
Examples
A forecast presented honestly
REVENUE FORECAST (next 2 mo)Point: 6.1M Range: 5.5M - 6.7M (80% interval)Assumptions: ordered pipeline converts at 82% velocity; no major price change; renewal band model-consistent.Validation (rolling): MAE 4.2% vs naive baseline 6.8%.Steer: funnel coverage < 1.0 would drag the range low — that's the dial we can move, not the forecast itself.A forecast is a range with assumptions and validation — a steering instrument, not a prophecy.
In production
- Monthly revenue forecast with an 80% band steers sales capacity planning.
- A churn-risk score drives the retention team's weekly outreach lists.
- Pipeline reconversion estimates fence quarter guidance early.
Best practices
- Always ship a range plus assumptions, never a naked point.
- Validate the model against a simple baseline before trusting it.
- Refresh on the cadence and log model drift.
- Label what business dials could move the forecast — prediction reachable by action.
Common mistakes
- Point-estimates pedestals — one number that 'will be' and everyone fights it.
- No assumptions line — the forecast reads as magic.
- Model never re-validated; drift tolerated until the boss is stunned.
- Prayers that predictions replace decisions — they're inputs, not commands.
Debugging tips
- Forecast off by repeated margins → seasonality/regime not in the model; re-specify.
- Range too wide to steer → interval honest but useless; tighten via better features.
- Execs read the point, ignore the band → render the range first, label it 'range of likely'.
Interview checkpoint
Hands-on lab
- Fit a simple linear trend to monthly history
- Compute point and interval for next month
- Compare against a naive baseline and label assumptions
Forecast a Trend in SQLite and Rule the Band
50 min · advanced
Quick quiz
A predictive BI forecast should ship with:
A forecast should replace the decision, not inform it.