Home › Glossary › Holt-Winters

Holt-Winters

Holt-Winters is a triple-exponential-smoothing method for forecasting time series with both trend and seasonality. Published by Charles Holt and Peter Winters in the early 1960s, it decomposes a series into three components — level, trend, and seasonality — each updated via its own smoothing parameter (α, β, γ). The forecast for horizon h is &Yhat; = (L + h×T) × St+h-m.

Dr Waqas Rafique Dr Waqas Rafique · Founder & CTO · PhD, Statistical Machine Learning
· About

When to use it

Time series with both trend and seasonality — e-commerce GMV, SaaS MRR, energy demand, agency monthly billings. At least two full seasonal cycles of history.

When not to use it

Pure white-noise series, series with structural breaks (relaunch, pricing change), or series with too few observations (<24 months for monthly data).

The three smoothing parameters

α controls how much the level reacts to new data (typically 0.1–0.5). β controls trend reactivity. γ controls seasonal reactivity. All bounded between 0 and 1. Auto-tuning via grid search over {0.1, 0.3, 0.5, 0.7, 0.9} works in practice.

Additive vs multiplicative seasonality

Multiplicative: the seasonal swing scales with the level (typical for revenue). Additive: the swing is constant in absolute terms (typical for temperature, headcount).

Alternatives to consider

Prophet (Facebook) for series with multiple seasonalities or known holidays. ARIMA for series with autocorrelation but no seasonality. State-space models (TBATS) for complex seasonal patterns. For most monthly business data, plain Holt-Winters wins on simplicity.

Related

Frequently asked questions

Is Holt-Winters the same as exponential smoothing?
Holt-Winters is *triple* exponential smoothing — it adds level, trend, and seasonality components. Single exponential smoothing has only level. Double (Holt) has level + trend. Holt-Winters is the version that handles seasonality.
Who invented Holt-Winters?
Charles Holt (level + trend, 1957) and his student Peter Winters (extending to seasonality, 1960). Both at Carnegie Institute of Technology, originally funded by the Office of Naval Research.
Is Holt-Winters still useful?
Yes — for monthly business data with seasonality, it remains one of the most reliable forecasting methods. Newer methods (Prophet, ARIMA, neural networks) have advantages on specific data types but rarely beat Holt-Winters by enough to justify their complexity for typical SaaS, e-commerce, or finance forecasting.