Feature Engineering

Compute statistical summaries over trailing time windows to capture recent supplier performance trends.

• Rolling Mean Transit Time: 7-day, 30-day, and 90-day averages of actual delivery durations
• Rolling Standard Deviation: Measures recent lead time variability for safety stock calculations
• Exponentially Weighted Moving Average (EWMA): Applies higher weight to recent observations, making the feature responsive to sudden shifts

Example: A 30-day rolling mean of 12.3 days vs. a 7-day rolling mean of 18.7 days signals rapid supplier deterioration.

Create features from past values of the target variable or related time series to capture autocorrelation patterns.

• Lag-1 Lead Time: The delivery duration of the immediately preceding order from the same supplier
• Lag-7 Seasonality: Lead time from 7 days prior to capture day-of-week effects
• Difference Features: Change in lead time between consecutive orders (lag-1 minus lag-2)

Critical for: Capturing momentum — if the last three shipments were late, the next one likely will be too.

Engineer features that encode historical supplier reliability patterns as static or slowly-changing covariates.

• On-Time In-Full (OTIF) Rate: Percentage of orders delivered complete and on schedule over the past quarter
• Mean Absolute Deviation (MAD): Average absolute difference between promised and actual delivery dates
• Bias Indicator: Signed metric showing whether a supplier systematically over-promises (negative bias) or under-promises (positive bias)
• Response Time Percentile: How quickly a supplier acknowledges purchase orders relative to peers

These features enable the model to learn supplier-specific correction factors.

Transform temporal metadata into features that capture cyclical patterns and known disruptions.

• Cyclical Encoding: Map day-of-week and month-of-year onto a unit circle using sine/cosine transforms to preserve periodicity
• Holiday Proximity: Days until or since major regional holidays (Chinese New Year, Golden Week, Christmas)
• Fiscal Period Flags: Binary indicators for month-end, quarter-end, and year-end shipping surges
• Business Day Counter: Number of actual working days between order and expected delivery, excluding weekends and holidays

Example: Lead times spike 300% during the week before Chinese New Year due to factory shutdowns.

Combine multiple raw variables to capture non-linear relationships that individual features miss.

• Distance × Carrier Type: Interaction term capturing that air freight speed advantage grows with distance
• Order Quantity ÷ Supplier Capacity: Ratio indicating supplier utilization pressure
• Season × Port: Categorical cross capturing that specific ports have unique seasonal congestion profiles
• Product Category × Origin Country: Encodes regulatory and customs complexity

Technique: Use feature crosses in gradient boosting models or explicit interaction terms in linear models to expose multiplicative effects.

Replace categorical variables with a numeric representation derived from the target variable, handling categories with many unique values.

• Supplier-Level Mean Lead Time: Replace supplier ID with their historical average delivery duration
• Lane-Level Variability: Encode origin-destination pairs with their standard deviation of transit time
• Carrier Reliability Score: Map carrier ID to their OTIF percentage

Regularization: Apply smoothing (Bayesian target encoding) or cross-fold computation to prevent overfitting on rare categories. A supplier with only 3 shipments should not get an extreme encoding.

A foundational technique that computes aggregate metrics over a moving time window to capture local trends and smooth out noise.

• Rolling Mean: Calculates the average transit time over the last 7, 14, or 30 days to establish a dynamic baseline.
• Rolling Standard Deviation: Measures recent lead time variability to quantify supplier inconsistency.
• Exponentially Weighted Moving Average (EWMA): Applies higher weights to recent observations, making the feature more responsive to sudden shifts in performance.

Example: A 14-day rolling average of port dwell times can signal emerging congestion before it becomes a systemic delay.

Creates input variables by shifting a time series backward by a specific number of periods, allowing a model to learn from past values.

• Autocorrelation Analysis: Use partial autocorrelation function (PACF) plots to identify statistically significant lags.
• Seasonal Lags: Capture weekly or monthly patterns by creating features lagged by 7 or 30 days.
• Cross-Correlation: Lag features from an external variable, such as shifting fuel prices by 3 days to predict carrier availability.

Example: A model predicting tomorrow's delivery time might use transit_time_lag_1 (yesterday's time) and transit_time_lag_7 (last week's time) as inputs.

Extracts categorical and cyclical components from timestamps to help models learn recurring temporal patterns.

• Categorical Features: Isolate day_of_week, month, quarter, and is_weekend to capture discrete behavioral shifts.
• Cyclical Encoding: Transform hour and day_of_year into sine and cosine coordinates to preserve their circular nature (e.g., 23:00 is close to 00:00).
• Holiday Proximity: Create a feature counting days until or since a major holiday to model demand surges and carrier capacity crunches.

Example: A sin_day_of_year and cos_day_of_year pair allows a model to understand that December and January are seasonally adjacent.

Engineers composite features by combining two or more categorical or numerical variables to capture non-linear, synergistic effects.

• Pairwise Combinations: Multiply supplier_lead_time_avg by carrier_reliability_score to create a risk-weighted lead time signal.
• Grouped Aggregations: Calculate the average On-Time In-Full (OTIF) rate for a specific supplier_id and origin_port pair.
• Ratio Features: Compute the ratio of actual_transit_time to contracted_transit_time to create a normalized performance index.

Example: A feature like avg_delay_for_supplier_on_route isolates a specific vendor's chronic performance issue on a particular lane.

Replaces categorical variables with a numeric value derived from the target variable, handling features like supplier_id that have thousands of unique levels.

• Mean Encoding: Replace each supplier with the historical mean of their actual lead time.
• Weight of Evidence (WoE): Calculate the log-odds of a late delivery for each category, commonly used in risk modeling.
• Smoothing & Cross-Validation: Apply additive smoothing to prevent overfitting on rare categories and use out-of-fold means to avoid data leakage.

Example: Instead of a one-hot encoded vector for 10,000 suppliers, use a single numeric column representing each supplier's average historical delay.

Creates binary or distance-based features that signal the occurrence and proximity of external shocks to a shipment's path.

• Binary Flags: A 1 if a geopolitical event, natural disaster, or labor strike is active on the shipment date, 0 otherwise.
• Geospatial Proximity: Calculate the Haversine distance between a shipment's planned route and an active disruption zone.
• Temporal Decay: Model the lingering effect of a disruption by creating a feature that decays exponentially from the event's end date.

Example: A port_congestion_flag feature activates when the vessel queue at the destination port exceeds the 90th historical percentile.