Orion

How Orion Works

Orion uses an ensemble of forecasting models, combining their predictions for robust accuracy across different demand patterns.

Ensemble Models

• Exponential Smoothing (ETS): Captures level, trend, and seasonality
• ARIMA: Handles autocorrelation in time series
• Prophet-style decomposition: Multiple seasonality patterns
• Gradient Boosting: Captures non-linear patterns and features

The ensemble weights are learned from your data, giving more weight to models that perform better for each product category.

Pattern Detection

Orion automatically detects and models:

• Trend: Long-term growth or decline
• Weekly Seasonality: Day-of-week patterns
• Monthly Seasonality: Start/end of month effects
• Annual Seasonality: Holiday and seasonal patterns
• Promotions: Lift from known promotional events

Confidence Intervals

Every forecast includes confidence intervals showing the range of likely outcomes:

• 50% CI: Likely range for typical planning
• 80% CI: Conservative range for inventory safety
• 95% CI: Wide range for risk assessment

Forecast Outputs

Inputs Required

Drift Detection

Orion continuously monitors forecast accuracy and detects when patterns change:

• Concept Drift: Underlying demand patterns changing
• Data Drift: Input data distribution changing
• Model Degradation: Accuracy declining over time

When drift is detected, Orion alerts you and automatically retrains affected models.

Integration with Other Agents

Orion's forecasts feed into other agents:

• → Nova: Forecasts drive safety stock and reorder calculations
• → Astra: Demand predictions inform price elasticity
• → Pulse: Baseline forecasts help measure promotion lift

Example Forecast

Best Practices

• Ensure at least 1 year of history for accurate seasonality
• Tag promotional periods in your data for better modeling
• Use 80% CI for safety stock calculations
• Review drift alerts promptly to maintain accuracy