Machine Learning
What is feature engineering?
Feature engineering is the work of turning raw event data into the numeric signals a model can actually learn from - and it is usually where most of a project's accuracy comes from.
Why it dominates model choice
Swapping XGBoost for LightGBM moves accuracy by a small margin. Adding one well-constructed feature - the ratio of this week's bet size to the player's own 90-day baseline - can move it far more. On tabular business problems, teams that spend their time on features beat teams that spend it on architecture, almost without exception.
The feature types that carry most signal
Rolling-window aggregates over 7, 30, and 90 days. Ratios of recent to baseline behaviour, computed per customer rather than against a global average. Cadence and gap features - time between events, and how that gap compares to the personal norm. Sequence features - direction of change, not just level. Level features almost always underperform change features.
The point-in-time rule
Every feature must be computable using only data that existed at the moment of prediction. This sounds obvious and is violated constantly, usually through a table that gets updated in place rather than appended. If your database overwrites a customer's status field, you cannot reconstruct what that field said last March, and any feature built on it leaks.
Build it once, use it twice
The pipeline that computes features for training must be the same code that computes them in production. Two separate implementations drift apart within weeks, and the resulting gap between offline evaluation and live performance is one of the hardest bugs to find in an ML system.
We build automation systems for small teams. Free audit call to map your specific workflows - no pitch, just a plan.