By the end of this lesson, the participant will be able to:
- understand what a predictive model is in practical terms
- recognize relevant use cases such as churn, subscription propensity, or recommendation
- understand basic criteria to evaluate whether a model is useful
- identify risks such as bias, overconfidence, or lack of interpretability
From hypothesis to prediction
In module 2, the participant learned to formulate hypotheses and think about basic experimentation. In this block, a new step appears: using data to anticipate probable behaviors.
The goal is not to replace human judgment with models, but to add an extra layer of help for better prioritization. A Data Driven organization starts to ask not only what happened and what could be tested, but also:
- which users are most likely to convert
- which profiles are at risk of abandonment
- which content is best to recommend based on history
What is a predictive model?
A predictive model uses observed signals to estimate the probability of a future behavior. In media, some typical cases may include:
- registration propensity
- subscription propensity
- churn risk
- probability of responding to a newsletter
- content recommendation
The key idea is not to master the mathematical technique, but to understand what problem it is trying to solve and how it can help make better decisions.
What makes a model useful?
A model is not useful because it is complex. It is useful if it helps make a decision better than before. To evaluate its usefulness, ask questions like:
- Does it solve a real problem?
- Is it actionable in time?
- Does it improve a decision that was previously worse or slower?
- Does it work reasonably well in different segments?
- Is it understandable to those who need to use it?
Risks and limitations
Models also have clear limits:
- they may amplify previous biases
- they may deteriorate over time if patterns change
- they may create a false sense of precision
- they may be poorly interpretable for decision-makers
For this reason, a mature Data Driven culture does not idolize models. It evaluates them, monitors them, and uses them cautiously.
AI applied to analytics
At this level, it is also important to introduce a basic idea about AI applied to analytics: automation and prediction can help a lot, but they must be inserted within a framework of supervision, validation, and responsibility.
Try it yourself
Your data team is proposing three predictive models. Here’s what they know so far:
Model X — Churn predictor
Flags subscribers showing a drop in visit frequency, lower newsletter engagement, and reduced recirculation over a 21-day window. Generates a daily list for the retention team to act on.
Model Y — Registration propensity scorer
Scores all anonymous users by likelihood to register, based on visits, thematic affinity, and time-on-site. Updates weekly. No specific action has been defined yet for high-scoring users.
Model Z — Content recommendation engine
Trained on 6 months of user behaviour from a period when the site had a very different editorial focus. Recommends related articles at the end of each piece. Has not been retrained since launch — 8 months ago.
Consider:
- Which model is most actionable right now — and why? Use the criteria from this lesson: does it solve a real problem, arrive in time, and improve a concrete decision?
- Model Y scores users but has no action attached. What problem does this create — and what would you add to make it useful?
- What specific risk does Model Z carry? What would you do before continuing to rely on it?
- Pick one of the three models and describe a realistic scenario where it could introduce bias into an editorial decision — and how you would detect it. (Example to get you started: what if the training data over-represented a particular type of user?)
A model is only as useful as the decision it improves. Precision without action is just noise.