SAM Forecasting Models: Complete Catalog

Overview

SAM Integration provides access to 12+ state-of-the-art forecasting algorithms, ranging from traditional statistical methods to cutting-edge neural networks. Our AI system automatically selects the optimal combination based on your data characteristics, ensuring maximum accuracy and reliability.

Model Categories

Statistical Models - Proven & Reliable

Traditional time series methods with decades of validation in business applications.

Neural Networks - Advanced & Adaptive

Modern deep learning approaches that excel with complex patterns and large datasets.

Specialized Models - Purpose-Built

Algorithms designed for specific use cases like seasonal business data or trend analysis.

Simple Models - Fast & Interpretable

Straightforward approaches ideal for baseline comparisons and quick insights.

---

Statistical Models

ARIMA (AutoRegressive Integrated Moving Average)

Best For: Data with clear trends, no seasonal patterns

• Strengths: Excellent trend modeling, statistical rigor, interpretable parameters
• Data Requirements: Minimum 50 observations, works with non-stationary data
• Processing Time: Medium (2-5 minutes for optimization)
• Use Cases: Revenue forecasting, economic indicators, non-seasonal business metrics

When to Use:

• Clear upward or downward trends
• No seasonal patterns (daily, weekly, monthly cycles)
• Need statistical significance testing
• Medium-sized datasets (50-1000 observations)

SARIMA (Seasonal ARIMA)

Best For: Data with both trends and seasonal patterns

• Strengths: Handles complex seasonality, robust trend modeling, statistical foundation
• Data Requirements: Minimum 100 observations, prefers multiple seasonal cycles
• Processing Time: High (5-15 minutes for optimization)
• Use Cases: Retail sales, seasonal demand, weekly/monthly business cycles

When to Use:

• Strong seasonal patterns (weekly, monthly, quarterly)
• Clear trends combined with seasonality
• Need detailed statistical analysis
• Sufficient historical data (2+ years)

Exponential Smoothing

Best For: Stable data with moderate seasonality, robust to outliers

• Strengths: Outlier resistant, handles missing data well, fast execution
• Data Requirements: Minimum 30 observations, works with sparse data
• Processing Time: Low (1-2 minutes)
• Use Cases: Inventory planning, stable product demand, operational metrics

When to Use:

• Data contains outliers or anomalies
• Missing values in historical data
• Need fast, reliable forecasts
• Stable business environment

Theta Model

Best For: Simple trend patterns, benchmark comparisons

• Strengths: Simple and fast, good baseline performance, minimal parameters
• Data Requirements: Minimum 20 observations
• Processing Time: Very Low (<1 minute)
• Use Cases: Quick forecasts, baseline comparisons, simple trend analysis

When to Use:

• Need rapid forecasting results
• Simple data patterns
• Benchmark against more complex models
• Limited computational resources

---

Neural Network Models

N-HiTS (Neural Hierarchical Interpolation for Time Series)

Best For: Large datasets, complex patterns, long-term forecasting

• Strengths: Excellent accuracy on large datasets, handles multiple seasonalities
• Data Requirements: Minimum 200 observations, benefits from GPU acceleration
• Processing Time: Medium-High (3-10 minutes with GPU)
• Use Cases: Demand forecasting, financial markets, large-scale operations

When to Use:

• Large historical datasets (200+ observations)
• Multiple seasonal patterns
• High accuracy requirements
• GPU resources available

TFT (Temporal Fusion Transformer)

Best For: Complex temporal patterns, multi-scale seasonality

• Strengths: State-of-the-art accuracy, attention mechanism, interpretability
• Data Requirements: Minimum 300 observations, GPU recommended
• Processing Time: High (5-20 minutes with GPU)
• Use Cases: Financial forecasting, complex business cycles, research applications

When to Use:

• Maximum accuracy requirements
• Complex, multi-scale patterns
• Need model interpretability
• Sufficient computational resources

GRU (Gated Recurrent Unit)

Best For: Sequential patterns, moderate computational requirements

• Strengths: Good balance of accuracy and speed, handles sequences well
• Data Requirements: Minimum 100 observations, GPU acceleration available
• Processing Time: Medium (2-8 minutes with GPU)
• Use Cases: Sales forecasting, user behavior, operational planning

When to Use:

• Sequential dependencies in data
• Balance between accuracy and speed
• Moderate dataset sizes
• Standard neural network applications

TCN (Temporal Convolutional Network)

Best For: Long-term dependencies, parallel processing

• Strengths: Fast training, captures long-term patterns, parallelizable
• Data Requirements: Minimum 150 observations, GPU acceleration beneficial
• Processing Time: Medium (2-6 minutes with GPU)
• Use Cases: Long-term planning, capacity forecasting, strategic analysis

When to Use:

• Long-term forecasting horizons
• Need fast neural network training
• Complex temporal dependencies
• Parallel processing capabilities

---

Specialized Models

Prophet (Facebook's Algorithm)

Best For: Business data with holidays, missing values, outliers

• Strengths: Robust to outliers, handles missing data, holiday effects
• Data Requirements: Minimum 100 observations, flexible with data quality
• Processing Time: Medium (2-5 minutes)
• Use Cases: Business metrics, user engagement, marketing analytics

When to Use:

• Business data with holiday effects
• Irregular data collection
• Need robust, reliable forecasts
• Data quality concerns

TBATS (Trigonometric, Box-Cox, ARMA, Trend, Seasonal)

Best For: Complex seasonality, multiple seasonal periods

• Strengths: Handles complex seasonality, automatic transformation selection
• Data Requirements: Minimum 200 observations, multiple seasonal cycles
• Processing Time: High (10-30 minutes)
• Use Cases: Complex seasonal business, multiple time cycles, detailed analysis

When to Use:

• Multiple seasonal patterns (daily + weekly + monthly)
• Complex business seasonality
• Advanced statistical modeling needs
• Sufficient processing time available

---

Simple Models

Moving Averages (4, 8, 13 weeks)

Best For: Baseline forecasts, trend smoothing, quick insights

• Strengths: Fast execution, easy interpretation, stable predictions
• Data Requirements: Minimum data equal to window size
• Processing Time: Very Low (<30 seconds)
• Use Cases: Baseline comparisons, trend analysis, quick estimates

When to Use:

• Need immediate results
• Simple trend analysis
• Baseline performance comparison
• Stable, predictable data

---

Model Selection Guide

Automatic Selection Criteria

Our AI system selects models based on these data characteristics:

For Seasonal Data (Strong Patterns)

1. SARIMA - Statistical rigor with seasonality
2. Prophet - Robust handling of business seasonality
3. TFT - Maximum accuracy for complex patterns
4. Exponential Smoothing - Fast, reliable seasonal modeling

For Trending Data (Growth/Decline)

1. ARIMA - Classic trend modeling
2. Prophet - Flexible trend handling
3. N-HiTS - Neural network trend capture
4. GRU - Sequential trend modeling

For Large Datasets (1000+ observations)

1. N-HiTS - Designed for large-scale data
2. TFT - Transformer architecture benefits
3. TCN - Parallel processing advantages
4. Prophet - Scalable performance

For Noisy/Outlier Data

1. Prophet - Robust to anomalies
2. Exponential Smoothing - Outlier resistant
3. GRU - Neural robustness
4. Moving Averages - Natural smoothing

For Fast Results (< 2 minutes)

1. Theta - Minimal processing time
2. Moving Averages - Instant results
3. Exponential Smoothing - Quick optimization
4. ARIMA - Fast convergence

Performance Matrix

Model	Accuracy	Speed	Complexity	Seasonality	Trend	Outlier Robust
ARIMA	High	Medium	Medium	❌	✅	❌
SARIMA	High	Low	High	✅	✅	❌
Exp Smoothing	Medium	High	Low	✅	✅	✅
Prophet	High	Medium	Medium	✅	✅	✅
N-HiTS	Very High	Medium	High	✅	✅	Medium
TFT	Very High	Low	Very High	✅	✅	Medium
GRU	High	Medium	High	Medium	✅	Medium
TCN	High	High	High	Medium	✅	Medium
Theta	Medium	Very High	Very Low	❌	✅	❌
Moving Avg	Low	Very High	Very Low	❌	Medium	✅

GPU Acceleration

Supported Models

Neural network models benefit significantly from GPU acceleration:

• N-HiTS: 3-5x faster training and inference
• TFT: 4-8x faster with complex architectures
• GRU: 2-4x faster with parallel processing
• TCN: 3-6x faster with convolutional operations

Performance Benefits

• Reduced Processing Time: Minutes instead of hours
• Larger Model Capacity: Handle more complex patterns
• Batch Processing: Multiple forecasts simultaneously
• Real-time Updates: Faster model retraining capabilities

---

How SAM Selects Models

Intelligent Model Selection Process

SAM Integration automatically chooses the best forecasting models for your data through a 3-step AI-driven process:

Step 1: Data Analysis

Our system analyzes your time series across 25+ characteristics:

• Seasonality: Detects weekly, monthly, quarterly patterns
• Trends: Identifies growth, decline, or stability
• Data Quality: Assesses completeness and outliers
• Volatility: Measures data stability and variability
• Size & Complexity: Evaluates dataset characteristics

Step 2: Model Scoring

Each of the 12+ available models receives a suitability score (0-10):

• Statistical Models (ARIMA, SARIMA): Best for clear trends and seasonal patterns
• Neural Networks (N-HiTS, TFT): Optimal for large, complex datasets
• Specialized Models (Prophet): Ideal for business data with holidays/outliers
• Simple Models (Moving Averages): Perfect for quick, stable forecasts

Step 3: Smart Selection

The AI doesn't just pick the highest scores - it ensures diversity:

• Balanced Portfolio: Combines different model types for robustness
• Optimal Count: Selects 2-5 models based on data complexity
• Performance Priority: Balances accuracy with processing speed
• Category Limits: Prevents over-reliance on any single approach

What You See

When forecasting starts, you'll receive:

• Selected Models: "AI chose Prophet, SARIMA, and N-HiTS"
• Selection Reason: "Best for seasonal business data with growth trends"
• Expected Accuracy: "Excellent performance anticipated"
• Processing Time: "Estimated completion in 8-12 minutes"

User Control Options

While AI selection is recommended, you can:

• Specify Models: Choose exact algorithms if needed
• Set Priorities: Emphasize speed vs accuracy
• Use Presets: Industry-optimized combinations available

---

Next Steps: Learn how to interpret your forecasting results in our Understanding Results guide, or explore the Technical Architecture to understand how these models work together at scale.