Feature Engineering as Leverage

Where the lift comes from

Past the baseline, feature engineering buys more performance per hour than hyperparameter tuning. The trick is engineering features that respect the time-of-prediction constraint. Aggregations over a time window, ratios, lags, and domain-specific groupings consistently outperform raw fields.

The pattern catalog

Time-window aggregates — count/sum/mean over the last 7/30/90 days, computed as of prediction time.
Ratios — value over a baseline ("this user's spend / their plan's average spend").
Recency / frequency — days since last action, count of actions per period.
Categorical interactions — pairs that represent business segments ("tier × region").
Cross-entity rollups — features about the user's organization, not just the user.

The leakage check on every feature

For each engineered feature, write a one-sentence proof that it could be computed at prediction time, with the data available at that exact moment. If the proof is hard, the feature is leakage.

Code

Time-aware rolling window in pandas·python

df = df.sort_values(["user_id", "event_time"])
df["tickets_30d"] = (
    df.groupby("user_id")
      .rolling("30D", on="event_time")["ticket_count"]
      .sum()
      .reset_index(level=0, drop=True)
)

Ratio and recency features·python

df["spend_vs_plan_avg"] = df["monthly_spend"] / df.groupby("plan_tier")["monthly_spend"].transform("median")
df["days_since_last_login"] = (df["prediction_time"] - df["last_login_at"]).dt.days

Where the lift comes from

The pattern catalog

The leakage check on every feature

Code

External links

Exercise

Progress

Comments 0