Predictive Analytics
Predictive analytics uses historical data and statistical models to forecast what happens next: revenue, churn, demand, which lead converts. It became standard equipment because every BI platform added a forecasting layer and every ad platform ships predicted audiences. A forecast now arrives with the dashboard. How much confidence that forecast deserves depends on what the model was trained on.
How it actually works
The off-the-shelf version fits a trend to internal history. The BI tool projects revenue from past revenue, the CRM scores leads on past closed-won patterns, the ad platform predicts converters from past converters. The models vary in sophistication; the training data is the same, the company's own records plus the platforms' reporting.
Forecasts built only on internal data and platform data attach a goal to last year's outcomes. A model like that extrapolates competently inside the world it was trained on. The forces that decide whether next year resembles last year, a competitor's move, a channel's economics shifting, a change in what customers are frustrated by, sit outside that training data, and the model returns the same confident number whether or not they are present.
A confident wrong forecast does more damage than no forecast, because budgets, hiring, and targets get set to it. The two directions of error carry different bills: an over-forecast costs the cost of capital on inventory built for demand that never arrived plus the stranded headcount hired to serve it, while an under-forecast costs the contribution margin on stockouts and turned-away demand. The two costs are rarely equal, so a forecast should state which direction of error is more expensive for this business and be held conservative in that direction. Growth gets planned against a curve that assumed the environment holds, and when the environment moves, the miss reads as an execution failure instead of a modeling gap. The correction belongs in the inputs: track the external forces alongside the internal history, use qualitative signals as leading indicators, and treat the model's output as a baseline to challenge, held conservative, with aggressive targets funded by evidence instead of extrapolation.
In practice
At Potbelly, the forecast was validated against a controlled result before it was trusted. Holding budget level, the measurement and modeling work showed spend could run five times more efficient. Only then did spend scale, and the business grew 30% and then 19% year over year. The forces outside the training data are what internal history cannot forecast. At an insurance brand spending $36 million a year on advertising, a lost lawsuit in one state forced compliance to sanitize the ad messaging, a volatile political climate raised category demand and masked the ads' falling effectiveness, and a denied customer claim became a national story just as the climate returned to baseline. Acquiring a customer became six times more expensive and LTV was cut in half in what appeared to be a matter of months, and no model trained on the company's own records carried any of those variables ([leading indicators](/insights/s-curve-of-growth#leading-indicators)).
Where we come in
We extract features from raw and qualitative data to train machine learning models, so the forecast reflects external forces and customer signals as well as internal history. Predictions are then validated with controlled experiments before budget moves. This method identified over $12 million in waste at USCCA.
Start a Revenue Health Pre-Assessment →See it in action
Related terms
- Forecasting
- Projecting a metric forward. The general practice predictive analytics automates, and the place where its assumptions do the most damage when hidden.
- Leading vs lagging indicators
- Lagging indicators, like revenue, report the past. Leading indicators, like qualitative demand signals, move first. Models trained only on lagging data project last year's pattern onto next year.
- Feature engineering
- Turning raw data into the inputs a model learns from. What gets engineered in decides what the model can see; external forces left out are absent from the forecast.
- Propensity model
- A model scoring how likely an individual is to act: buy, churn, click. Behind lead scoring and predicted audiences, and trained on the same internal history as everything else.
- Marketing mix modeling (MMM)
- Estimating each channel's revenue contribution from historical data at the portfolio level. A statistical sibling of predictive analytics used for budget allocation rather than forecasting.
- Overfitting
- A model that learned its training data too specifically and fails on anything new. The formal version of mistaking last year's pattern for a law.