Intelligent Demand Forecasting System

The Challenge

Supply chain forecasting is notoriously difficult because not all products behave the same way. A best-seller with consistent daily sales needs a different forecasting approach than a slow-moving spare part ordered once a month.

The existing system used a one-size-fits-all approach—applying the same simple moving average to all 10,000+ SKUs. This led to:

• Frequent stockouts on fast-moving items (losing sales)
• Excess inventory on slow-movers (tying up capital)
• Manual overrides by planners consuming hours daily
• No visibility into why forecasts were off

The Solution

I built an intelligent forecasting system that classifies each SKU's demand pattern and automatically selects the most appropriate algorithm from an ensemble of 15+ models.

Pattern Classification (ADI/CV²)

The system uses two key metrics to classify demand patterns:

• ADI (Average Demand Interval) — How often does demand occur?
• CV² (Coefficient of Variation Squared) — How variable is demand size?

Algorithm Ensemble

The system includes 15+ forecasting algorithms, each optimized for different demand characteristics:

+ Moving Averages, Linear Regression, LSTM, Theta Method, Naive Methods, and more...

Implementation

The pattern classifier uses a clean, configurable design:

class DemandPatternClassifier:
    """Classify SKU demand patterns using ADI/CV² analysis."""
    
    def __init__(self, adi_threshold=1.32, cv2_threshold=0.49):
        self.adi_threshold = adi_threshold
        self.cv2_threshold = cv2_threshold
        
    def classify(self, demand_series: pd.Series) -> DemandPattern:
        # Calculate Average Demand Interval
        non_zero_periods = (demand_series > 0).sum()
        adi = len(demand_series) / non_zero_periods if non_zero_periods > 0 else float('inf')
        
        # Calculate CV² (squared coefficient of variation)
        non_zero_demand = demand_series[demand_series > 0]
        cv2 = (non_zero_demand.std() / non_zero_demand.mean()) ** 2
        
        # Classify based on thresholds
        if adi < self.adi_threshold and cv2 < self.cv2_threshold:
            return DemandPattern.SMOOTH
        elif adi < self.adi_threshold and cv2 >= self.cv2_threshold:
            return DemandPattern.ERRATIC
        elif adi >= self.adi_threshold and cv2 < self.cv2_threshold:
            return DemandPattern.INTERMITTENT
        else:
            return DemandPattern.LUMPY

Tech Stack

Key Learnings