AI Forecasting Overview

The AI forecasting system uses multiple advanced machine learning models to predict future demand based on historical data. The system automatically selects the best performing model for your specific data patterns.

System Requirements

For reliable forecasts, you need:

• Minimum 6 months of historical data
• Daily sales records
• Regular data updates
• Clean data without large gaps

Understanding The Models In Our System

Traditional ML Models

1. Exponential Smoothing (Holt-Winters)

   • Pure statistical ML approach
   • No deep learning or AI components
   • Based on weighted averages and decay functions
   • Best for: Short-term forecasts with clear patterns

2. AutoARIMA

   • Classical statistical ML model
   • No neural networks or deep learning
   • Uses traditional time series decomposition
   • Best for: Medium-term forecasts with linear trends

3. Croston's Method

   • Specialized for intermittent demand
   • Uses alpha=0.1 smoothing parameter
   • Best for: Intermittent and lumpy demand patterns

4. Naive Seasonal

   • 7-day repeating pattern
   • Uses last week's pattern
   • Best for: Simple weekly patterns and benchmarking

AI-Enhanced Models

1. Prophet (Facebook/Meta)

   • Hybrid AI/ML approach
   • Uses modern Bayesian modeling
   • Not an LLM, but incorporates AI principles for:
     • Automatic changepoint detection
     • Pattern recognition
     • Anomaly detection
   • Best for: Complex seasonal patterns and long-term trends

2. NeuralProphet

   • Deep learning extension of Prophet
   • Combines neural networks with Prophet's architecture
   • Key features:
     • AR-Net for autocorrelation modeling
     • FFT for seasonal patterns
     • Deep learning components for non-linear relationships
   • Best for: Complex patterns with multiple seasonalities
   • Note: Requires more historical data than traditional Prophet

Important Note on AI/ML Terminology

• None of our models are Large Language Models (LLMs)
• We use specialized time series models, not general-purpose AI
• The "AI" components refer to advanced pattern recognition and automated optimization
• All models are specifically designed for numerical forecasting

How It Works

Automated Process

The system automatically:

1. Splits your historical data (80% training, 20% validation)
2. Trains all models on your historical data
3. Evaluates performance using MAPE and RMSE
4. Selects the best performing model for your final forecast

Reading Your Forecast

Forecast Sections

When you view a forecast, you'll see:

1. Daily Predictions

   • Predicted sales quantities and revenue
   • Weekday vs weekend patterns

2. Monthly Overview

   • Total expected sales
   • Average daily demand
   • Number of active selling days
   • Month-over-month comparison

Visual Indicators

• 🟦 Blue: Historical data
• 🟩 Green: Predictions
• ⚠️ Yellow: Unusual patterns or potential issues

Best Practices

When to Update Forecasts

• After significant sales events
• When adding new products
• Monthly for regular business planning
• Before making large inventory decisions

Tips for Better Results

1. Maintain Data Quality

   • Keep historical data current
   • Review forecasts monthly
   • Document special events or promotions

2. Plan for Seasonality

   • Mark known busy periods
   • Plan for holiday seasons
   • Account for yearly patterns

Technical Details

Accuracy Metrics

• MAPE (Mean Absolute Percentage Error)
  • Lower is better
  • Example: 15% means predictions are off by 15% on average
• RMSE (Root Mean Square Error)
  • Shows error in actual units
  • Matches your sales data units

Simple Average Fallback

When insufficient historical data exists for advanced forecasting models (less than 2 data points), the system falls back to a simple average calculation:

• Uses your store's default forecasting days setting
• Calculates the average daily demand over this period
• Applies this average to future predictions

When Simple Averages Are Used

• New products with limited history
• Products with gaps in sales data
• Items with fewer than 2 data points
• Irregular or intermittent demand patterns

Tip: Adjust your store's default forecasting days based on your business needs. Consider:

• Product lifecycle
• Seasonal patterns
• Ordering frequencies
• Data reliability

Demand Pattern Classification

The system automatically classifies demand patterns using two key metrics:

• ADI (Average Demand Interval): Measures how frequently demand occurs
• CV² (Squared Coefficient of Variation): Measures demand variability

Classification Thresholds

• ADI threshold: 1.32
• CV² threshold: 0.49

Demand Pattern Types

1. Smooth (ADI < 1.32, CV² < 0.49)

   • Regular, consistent demand
   • Most common for fast-moving consumer goods
   • Optimal for all forecasting models
   • Prophet Parameters:
     • Higher seasonality sensitivity
     • Moderate changepoint detection

2. Intermittent (ADI ≥ 1.32, CV² < 0.49)

   • Irregular demand intervals
   • Consistent quantities when demand occurs
   • Prophet Parameters:
     • Lower seasonality sensitivity
     • Conservative changepoint detection

3. Erratic (ADI < 1.32, CV² ≥ 0.49)

   • Regular demand intervals
   • Highly variable quantities
   • Prophet Parameters:
     • Moderate seasonality sensitivity
     • Aggressive changepoint detection

4. Lumpy (ADI ≥ 1.32, CV² ≥ 0.49)

   • Irregular demand intervals
   • Highly variable quantities
   • Most challenging to forecast
   • Prophet Parameters:
     • Minimal seasonality sensitivity
     • Very conservative changepoint detection

5. Zero Demand

   • No historical demand
   • Falls back to conservative parameters
   • Typically uses simple average forecasting

Model Parameter Optimization

The system automatically adjusts forecasting parameters based on the identified demand pattern:

• Changepoint Detection: Controls how sensitive the model is to trend changes
• Seasonality Sensitivity: Adjusts how strongly seasonal patterns are weighted
• Seasonality Mode: Uses multiplicative seasonality for all patterns

Forecast Horizons

• Daily forecasts: First 30 days
• Weekly averages: Days 31-365
• Maximum forecast period: 365 days
• Confidence intervals: Exponentially increasing uncertainty
  • Base uncertainty: 30% of historical standard deviation
  • Grows exponentially with forecast horizon

Troubleshooting

If you notice forecast issues:

1. Verify historical data completeness
2. Check for unusual patterns or spikes
3. Contact support for interpretation help

Data Requirements

• Minimum 2 non-zero data points
• Up to 18 months of historical data used
• Daily frequency (gaps automatically filled with zeros)
• For reliable forecasts, recommended:
  • At least 120 days of data for cross-validation
  • Regular data updates
  • Clean data without large gaps