Feature Engineering

Feature engineering is the single most impactful step in predictive modeling. A mediocre model with excellent features outperforms an excellent model with mediocre features. Every time. The model is the brain. The features are the senses. A genius brain with poor eyesight will lose to an average brain with perfect vision.

A feature is any input signal the model uses to make predictions. Raw data — deal amount, days since last contact, company size — is Level 1. Engineered features — deal amount relative to average for that segment, contact frequency trend over last 30 days, company growth rate — is Level 2. Level 2 features carry more predictive signal because they encode relationships and context that raw numbers cannot.

1. Category 1: Behavioral Features Engineered from activity patterns. Not "emails sent" but "email velocity change over last 14 days." Not "meetings held" but "meeting frequency relative to deal stage average." Behavioral features capture momentum and trajectory, which predict outcomes better than static counts.
2. Category 2: Relative Features Engineered by comparing against a benchmark. Not "deal size: $85K" but "deal size: 1.7x segment average." Not "cycle length: 45 days" but "cycle length: 0.8x average for deals this size." Relative features normalize for segment differences and reveal whether a deal is above or below expected performance.
3. Category 3: Temporal Features Engineered from time-based patterns. Day of week, month, quarter, days until quarter-end, days since last interaction, time between stage transitions. Temporal features capture urgency signals that static data misses.
4. Category 4: Interaction Features Engineered by combining two features. "Deal size x number of stakeholders" captures complexity. "Win rate x deal age" captures aging risk. Interaction features model relationships between variables that single features cannot represent.

AI assists feature engineering by analyzing raw data and suggesting transformations. Feed a dataset to AI with the prompt: "Suggest 20 engineered features that might predict [outcome] from this data. For each, explain the hypothesis for why it would be predictive." The AI generates candidates. You validate them against historical outcomes. CIPHER's rule: a feature earns its place only if it improves prediction accuracy by at least 1% on the validation set.