Skip to main content

SAM Forecasting Methodology: How It Works

Overview​

SAM's Uni-Variate Forecasting employs a sophisticated 4-phase methodology that combines advanced statistical analysis, artificial intelligence, and enterprise-grade processing to deliver highly accurate, automated forecasts.

1. Intelligent Dataset Analysis​

Comprehensive Data Profiling​

Our system automatically analyzes your time series across 25+ statistical dimensions to understand the underlying patterns and characteristics:

Statistical Characteristics​

  • Central Tendency: Mean, median, mode analysis
  • Variability: Standard deviation, coefficient of variation
  • Distribution: Skewness, kurtosis, normality assessment
  • Data Quality: Missing values, zero counts, sparsity analysis

Time Series Properties​

  • Stationarity Testing: Augmented Dickey-Fuller test to determine if data needs differencing
  • Seasonality Detection: Multi-period analysis (52, 26, 12, 4 weeks) with strength measurement
  • Trend Analysis: Linear regression slope calculation with direction and magnitude
  • Residual Analysis: Error pattern identification and strength assessment

Data Complexity Assessment​

  • Outlier Detection: IQR-based anomaly identification with percentage calculation
  • Volatility Analysis: Coefficient of variation for stability assessment
  • Size Evaluation: Large vs small dataset determination for algorithm selection
  • Sparsity Measurement: Zero-value frequency for model suitability

Advanced Pattern Recognition​

Example Analysis Results:
• Seasonality Strength: 0.65 (Strong seasonal pattern detected)
• Trend Direction: Increasing (3.2% monthly growth)
• Stationarity: Non-stationary (requires differencing)
• Data Quality: 98.5% complete, 2.3% outliers
• Volatility: Moderate (CV = 0.45)

2. AI-Powered Model Selection​

Intelligent Scoring Algorithm​

Each available forecasting model receives a suitability score (0-10) based on dataset characteristics:

Model-Specific Evaluation Criteria​

  • Data Size Requirements: Minimum observations needed for reliable results
  • Stationarity Preferences: Whether model handles non-stationary data effectively
  • Seasonality Capabilities: Ability to capture and forecast seasonal patterns
  • Trend Handling: Effectiveness with increasing/decreasing/stable trends
  • Outlier Robustness: Performance degradation with anomalous data points
  • Computational Complexity: Processing time vs accuracy trade-offs

Smart Selection Process​

Step 1: Suitability Scoring​

Example Model Scores:
• SARIMA: 8.5/10 (High seasonality + trend handling)
• Prophet: 8.2/10 (Robust to outliers + flexible seasonality)
• N-HiTS: 7.8/10 (Large dataset + neural network advantages)
• ARIMA: 6.5/10 (Good trend handling, no seasonality)
• Exp Smoothing: 7.2/10 (Balanced performance + speed)

Step 2: Diversity Optimization​

Our system ensures balanced model selection across different categories:

  • Statistical Models: ARIMA, SARIMA, Exponential Smoothing
  • Neural Networks: N-HiTS, TFT, GRU, TCN
  • Advanced Models: Prophet, TBATS
  • Simple Models: Moving Averages, Theta

Step 3: Adaptive Selection​

The number of models selected adapts to dataset characteristics:

  • Small Datasets (1-2 categories): 2-3 high-quality models
  • Medium Datasets (3-5 categories): 3-4 diverse models
  • Large Datasets (5+ categories): 4-5 comprehensive models

Real-Time Performance Profiling​

Before full execution, our system estimates runtime for each selected model:

  • Sample-Based Profiling: Test execution on 10% of data
  • Scaling Calculations: Extrapolation to full dataset size
  • Resource Optimization: GPU vs CPU allocation decisions

3. Advanced Model Processing​

Hyperparameter Optimization​

Each model undergoes automated tuning using the Optuna framework:

ARIMA/SARIMA Models​

  • Parameter Space: p (0-5), d (0-2), q (0-5) combinations
  • Optimization Trials: 50 iterations with 5-minute timeout
  • Selection Criteria: AIC minimization for statistical significance
  • Validation Method: In-sample fit quality assessment

Neural Network Models​

  • Architecture Tuning: Hidden layer sizes, dropout rates, learning rates
  • Training Optimization: Early stopping, batch size adaptation
  • GPU Acceleration: CUDA utilization for faster computation
  • Cross-Validation: Time series split validation for robustness

Prophet Models​

  • Seasonality Components: Weekly, yearly pattern strength
  • Trend Flexibility: Changepoint detection sensitivity
  • Holiday Effects: Automatic holiday impact inclusion
  • Uncertainty Intervals: Bayesian posterior sampling

Parallel Execution Engine​

Multiple models execute simultaneously using advanced processing:

Background Processing​

  • Thread-Safe Operations: Non-blocking API responses
  • Status Tracking: Real-time progress monitoring
  • Error Handling: Graceful failure recovery with fallback options
  • Resource Management: Dynamic CPU/GPU allocation

Quality Assurance​

  • Validation Splits: 95% training, 5% testing for accuracy assessment
  • Convergence Checking: Model stability verification
  • Result Validation: Output sanity checks and boundary validation

4. Comprehensive Result Generation​

Multi-Level Validation​

Each model produces three types of predictions:

  1. Fitted Values: Historical period predictions (training data)
  2. Out-Sample Validation: Test period accuracy assessment
  3. Future Forecasts: User-specified horizon predictions

Advanced Metrics Calculation​

Accuracy Metrics​

  • RMSE (Root Mean Square Error): Overall prediction accuracy
  • MAPE (Mean Absolute Percentage Error): Percentage-based error measurement
  • Reliability Score: Confidence-adjusted accuracy (0-100 scale)
  • Accuracy Grade: Simplified rating (Excellent/Good/Fair/Poor)

Business Intelligence Metrics​

  • Growth Analysis: Historical vs forecast percentage changes
  • Trend Direction: Increasing/Decreasing/Stable classification
  • SPYA Comparisons: Same Period Year Ago analysis for seasonality
  • Forecast Stability: Consistency measurement across prediction horizon

Confidence Assessment​

  • Confidence Levels: High/Medium/Low reliability classification
  • Error Coefficients: Statistical uncertainty quantification
  • Forecast Ranges: Upper and lower prediction bounds

Multi-Format Output Generation​

Standardized Data Export​

9-column CSV format with complete forecast details:

Week | Week_Ending_Date | Product_Category | Forecast_Model | 
Actual_Values | Forecasted_Values | Root_Mean_Square_Error | 
Absolute_Error | Cumulative_Absolute_Error

Visual Analytics​

  • Interactive Charts: Actual vs predicted with error visualization
  • Model Comparisons: Side-by-side performance analysis
  • Trend Visualization: Long-term pattern identification
  • Confidence Bands: Uncertainty representation

Executive Reporting​

  • PDF Summary: Professional multi-page report with model rankings
  • Performance Dashboard: Key metrics visualization
  • Business Insights: Growth projections and trend analysis
  • Recommendation Engine: Best model identification with rationale

Quality Assurance & Validation​

Automated Quality Checks​

  • Data Integrity: Missing value handling, outlier treatment
  • Model Convergence: Training stability verification
  • Result Validation: Output range and trend reasonableness
  • Performance Benchmarks: Historical accuracy tracking

Error Handling & Recovery​

  • Graceful Degradation: Fallback to alternative models if primary fails
  • Partial Results: Delivery of available forecasts even with some model failures
  • Status Transparency: Clear communication of any processing issues
  • Recovery Options: Automatic retry mechanisms for transient failures

5. AI-Powered Business Intelligence​

Revolutionary Integration: SAM combines forecasting accuracy with GPT-4 intelligence to deliver not just predictions, but strategic insights, executive summaries, and actionable business recommendations.

Why AI Integration Matters​

  • Technical Translation: Statistical metrics become clear business insights
  • Strategic Context: Forecasts connected to business implications
  • Executive Communication: Results formatted for leadership consumption
  • Actionable Guidance: Specific recommendations for operations and strategy
  • Risk Intelligence: Automated uncertainty analysis with business context

Azure OpenAI Integration​

Enterprise-Grade AI Partnership​

  • Enterprise Security: Business-grade data protection and compliance
  • Scalable Performance: Multiple simultaneous analyses
  • Consistent Quality: Professional-grade content generation
  • Cost Optimization: Efficient token usage and intelligent caching

AI Processing Pipeline​

Forecast Results + Model Metrics + Business Context
           ↓
    Data Contextualization
           ↓
  Business Intelligence Generation
           ↓
     Azure OpenAI GPT-4
           ↓
Professional Business Intelligence Output

Methodology Advantages​

Scientific Rigor​

  • Multi-Model Ensemble: Reduces single-algorithm bias
  • Cross-Validation: Robust out-of-sample testing
  • Statistical Significance: AIC/BIC-based model selection
  • Uncertainty Quantification: Confidence interval generation

Enterprise Scalability​

  • Parallel Processing: Simultaneous multi-model execution
  • Resource Optimization: Dynamic GPU/CPU allocation
  • Background Operation: Non-blocking user experience
  • Cloud Integration: Unlimited storage and processing capacity

Business Intelligence​

  • Automated Insights: No manual interpretation required
  • Actionable Metrics: Direct business decision support
  • Risk Assessment: Confidence levels for uncertainty management
  • Trend Analysis: Strategic planning support

Next Steps: Explore our Available Models to understand which algorithms work best for your specific use case, or dive into Understanding Results to maximize the value of your forecasting outputs.