Engineer Time-Based Features for ML Models

This skill covers time-based feature engineering - extracting meaningful temporal patterns from datetime data to improve ML model performance, transforming raw timestamps into features that capture seasonality, trends, cyclical patterns, and temporal relationships. Calendar features are extracted via a function pulling year/month/day/hour/minute/day-of-week/day-of-year/week-of-year/quarter plus boolean indicators (is_weekend, is_month/quarter/year start/end). Cyclical encoding transforms periodic components into sine/cosine pairs so temporal continuity is preserved:

def create_cyclical_features(df, datetime_col):
    """Create cyclical encodings for periodic time features"""
    dt = pd.to_datetime(df[datetime_col])
    
    cyclical_features = pd.DataFrame({
        # Hour cyclical (24-hour cycle)
        'hour_sin': np.sin(2 * np.pi * dt.dt.hour / 24),
        'hour_cos': np.cos(2 * np.pi * dt.dt.hour / 24),
        
        # Day of week cyclical (7-day cycle)
        'dayofweek_sin': np.sin(2 * np.pi * dt.dt.dayofweek / 7),
        'dayofweek_cos': np.cos(2 * np.pi * dt.dt.dayofweek / 7),
        
        # Month cyclical (12-month cycle)
        'month_sin': np.sin(2 * np.pi * dt.dt.month / 12),
        'month_cos': np.cos(2 * np.pi * dt.dt.month / 12),
        
        # Day of year cyclical (365-day cycle)
        'dayofyear_sin': np.sin(2 * np.pi * dt.dt.dayofyear / 365.25),
        'dayofyear_cos': np.cos(2 * np.pi * dt.dt.dayofyear / 365.25),
    })
    
    return cyclical_features

Lag and window features cover historical lookback values at configurable lags (with proper per-entity grouping for panel data like per-customer or per-product series) and rolling mean/std/min/max over configurable window sizes, also entity-aware. Advanced temporal patterns cover time-since-event features (days elapsed since the last occurrence of a named event, forward-filled, plus boolean "event in last N days" flags) and temporal aggregations (weekly/monthly per-entity means with week-over-week and month-over-month percentage change, or simple weekly/monthly means and deviation-from-mean ratios when no entity grouping is used).

Best practices cover feature selection and validation (time-based cross-validation to prevent leakage, testing feature stability across periods, monitoring importance drift, removing highly correlated temporal features), memory and performance optimization (appropriate dtypes like int8 for booleans, numba-accelerated rolling calculations for large datasets, caching expensive computations, categorical encoding for high-cardinality time features), and domain-specific considerations (business-hours vs. after-hours patterns, holiday/special-event indicators, seasonal adjustment for trending variables, timezone handling for global datasets). A comprehensive pipeline function chains calendar features, cyclical features, and - when a target column is supplied - lag and rolling features into one combined DataFrame.

Use it when engineering datetime-derived features for a forecasting or predictive model - calendar/cyclical encodings, lag and rolling-window features, time-since-event features, or period-over-period aggregations. It is not a full time-series forecasting-model guide (ARIMA, Prophet, etc.) - it covers the feature layer built from timestamps, not the forecasting model itself.

Given a DataFrame with a datetime column - and optionally a target column and an entity column for grouped/panel data - it returns a features DataFrame combining calendar, cyclical, lag, rolling, time-since, and aggregation features ready to feed into a model.

Built on pandas (the .dt accessor, groupby/rolling/shift/pct_change) and numpy for cyclical trigonometric transforms, with numba referenced for accelerating rolling calculations on large datasets.

ML engineers and data scientists engineering time-based features for forecasting, churn, or any model where temporal patterns matter.