Skip to content

From Projections to Features

Analytical projections give you structured, queryable datasets, but AI models do not train on projections directly. They require features – measurable variables that capture meaningful aspects of the domain and provide signals for prediction.

Turning projections into features is where domain understanding meets the technical craft of feature engineering.

Why Features Define Model Success

The choice and quality of features often have a greater impact on model performance than algorithm selection.

A well-designed feature that reflects a genuine business signal can transform a model from mediocre to outstanding, while a misleading or unstable feature can undermine even the most advanced algorithm.

Good features are:

  • Relevant – they represent something that truly influences the target outcome
  • Consistent – they mean the same thing across the entire dataset
  • Explainable – domain experts can understand their meaning and origin

Event Sourcing and Feature Engineering

Event Sourcing offers a significant advantage for feature engineering: it gives you a complete, ordered history of what happened in the domain.

Because you can replay events for any point in time, features can be:

  • Historically accurate
  • Reproducible
  • Point-in-time correct

In our example of the library domain this means:

  • Derive a punctuality rate for each member from a projection of all loans
  • Calculate seasonal demand scores from time-series projections
  • Compute average reading times per genre from sequences of loan and return events

Each feature is grounded in real, immutable facts, making it easier to explain and validate.

The Role of Domain Understanding

Feature engineering is not just about transforming numbers – it depends on business context. Two features may look similar statistically, but one could be a powerful predictor while the other is just noise.

Domain experts can:

  • Identify which signals are likely to matter
  • Help interpret what they mean
  • Ensure features align with real-world processes

This link between features and real-world meaning is also essential for explainable AI, where every input value can be traced back to the events that produced it.

Designing for Flexibility

A feature that is not relevant today may become critical tomorrow. With Event Sourcing, you can always:

  • Replay historical events to produce new features
  • Iterate on model design without collecting new data
  • Experiment safely while preserving reproducibility

This flexibility supports iterative development and the long-term evolution of your AI capabilities.

Next up: Packaging Data Products – learn how to make your AI-ready datasets discoverable, reliable, and reusable across your organization.