asgard-ai-platform Exponential Smoothing
Overview
Exponential smoothing assigns exponentially decreasing weights to past observations. Three variants: Simple (SES, level only), Holt (level + trend), Holt-Winters (level + trend + seasonality). ETS framework (Error-Trend-Seasonality) provides a unified statistical model. Fast, interpretable, and competitive with complex models for short horizons.
When to Use
Trigger conditions:
- Quick forecasting with minimal configuration
- Short-horizon forecasts (1-2 seasonal cycles ahead)
- Data with clear level, trend, and/or seasonal components
When NOT to use:
- For long-range forecasts (uncertainty accumulates too fast)
- When external regressors are important (use regression or ML models)
Algorithm
IRON LAW: Smoothing Parameters Control the Bias-Variance Trade-Off
α (level), β (trend), γ (seasonality) range [0,1].
- α near 1: react quickly to changes, noisy forecasts (high variance)
- α near 0: smooth forecasts, slow to adapt (high bias)
Optimize via minimizing MSE on training data (or use information criteria).
Never hand-pick smoothing parameters without validation.
Phase 1: Input Validation
Identify components: level only (SES), level+trend (Holt), level+trend+seasonality (Holt-Winters). Determine: additive vs multiplicative trend/seasonality. Gate: Component structure identified, seasonal period known.
Phase 2: Core Algorithm
Holt-Winters (additive):
- Initialize: level₀ = mean(first season), trend₀ = (mean(season 2) - mean(season 1))/s, seasonal₀ from first season deviations
- Update equations at each t:
- Level: ℓₜ = α(yₜ - sₜ₋ₛ) + (1-α)(ℓₜ₋₁ + bₜ₋₁)
- Trend: bₜ = β(ℓₜ - ℓₜ₋₁) + (1-β)bₜ₋₁
- Seasonal: sₜ = γ(yₜ - ℓₜ) + (1-γ)sₜ₋ₛ
- Forecast: ŷₜ₊ₕ = ℓₜ + h×bₜ + sₜ₊ₕ₋ₛ
Phase 3: Verification
Check: in-sample RMSE, residual patterns. Compare against naive baselines (last value, seasonal naive). Gate: Beats naive baseline, residuals show no systematic pattern.
Phase 4: Output
Return forecasts with smoothed components.
Output Format
{
"forecasts": [{"period": "2025-04", "forecast": 1150, "level": 1100, "trend": 20, "seasonal": 30}],
"parameters": {"alpha": 0.3, "beta": 0.1, "gamma": 0.15},
"metadata": {"method": "holt_winters_additive", "seasonal_period": 12, "rmse": 45}
}
Examples
Sample I/O
Input: 36 months of monthly sales, clear upward trend, December spike Expected: Holt-Winters additive. Forecast continues trend with repeated December seasonality.
Edge Cases
| Input | Expected | Why |
|---|---|---|
| No trend, no seasonality | SES (α only) | Simplest variant suffices |
| Seasonal amplitude grows | Use multiplicative | Additive would underestimate peaks |
| Very short series (<2 seasons) | SES or Holt only | Can't estimate seasonality |
Gotchas
- Additive vs multiplicative: If seasonal swings grow proportionally with level, use multiplicative. Wrong choice produces poor forecasts, especially at extremes.
- Initialization sensitivity: The first season's values set the baseline. Poor initialization from noisy early data propagates through the entire forecast.
- Damped trend: For long horizons, linear trend extrapolation is unrealistic. Use damped trend (φ parameter) to flatten the trend over time.
- Multiple seasonalities: Standard Holt-Winters handles one seasonal period. For daily data with weekly AND yearly patterns, use TBATS or STL+ETS.
- Outlier sensitivity: A single outlier can shift the level estimate significantly (especially with high α). Pre-detect and handle outliers.
References
- For ETS framework and model selection, see
references/ets-framework.md - For damped trend variants, see
references/damped-trend.md