Exponential smoothing & Holt-Winters

No — not always. The exponential smoothing family, also called ETS (Error-Trend-Seasonal), is a collection of models that are fast to fit, easy to interpret, and surprisingly competitive with ARIMA on real-world data. The core idea is beautifully simple: weight observations so that the most recent value matters most and every older value matters a little less, with the influence decaying exponentially into the past.

Simple Exponential Smoothing (SES)

Simple Exponential Smoothing produces a forecast by maintaining a single running estimate called the level, updated after every new observation.

The recursive update rule is:

level_t = alpha * y_t + (1 - alpha) * level_(t-1)

• y_t is the observed value at time t.
• level_(t-1) is the level estimate from the previous step — a smoothed summary of all earlier history.
• alpha (the smoothing parameter) lives in the interval (0, 1) and controls how fast the memory fades.

When alpha is close to 1, the level lunges toward each new observation — the model is reactive but noisy. When alpha is close to 0, the level barely moves — the model is stable but slow to track genuine shifts. The one-step-ahead forecast is simply the current level: forecast_(t+1) = level_t.

Why this is genuinely exponential

Unrolling the recursion shows what is happening under the hood. The current level is:

level_t = alpha * y_t + alpha*(1-alpha)*y_(t-1) + alpha*(1-alpha)^2*y_(t-2) + ...

Every additional lag picks up another factor of (1-alpha). Because (1-alpha) is less than 1, those factors shrink toward zero — an exponential decay of weights as you look further into the past. The diagram below shows this visually.

Holt’s method — adding a trend

Holt’s method (also called double exponential smoothing) extends SES by tracking two quantities:

• Level l_t — the smoothed baseline, updated with smoothing parameter alpha.
• Trend b_t — the smoothed slope (rate of change), updated with smoothing parameter beta.

The forecast h steps ahead is l_t + h * b_t: start from the current level and project along the current trend. Beta, like alpha, lives in (0, 1) and controls how quickly the model revises its trend estimate.

Holt-Winters — adding seasonality

Holt-Winters (also called triple exponential smoothing) adds a third set of smoothed quantities: seasonal indices s_t, one for each period within the season (twelve values for monthly data with annual seasonality, seven for daily data with weekly seasonality).

A third smoothing parameter gamma controls how fast the seasonal indices adapt. The seasonal_periods parameter tells the model the length of one seasonal cycle.

There are two variants:

• Additive — the seasonal component is added to the level. Use this when the size of seasonal swings stays roughly constant regardless of the level of the series.
• Multiplicative — the seasonal component multiplies the level. Use this when seasonal swings grow proportionally as the level rises (common in retail and tourism data).

Together the three parameters — alpha for level, beta for trend, gamma for seasonality — define what the forecasting community calls the ETS (Error-Trend-Seasonal) model family. The “E” refers to the error structure (additive or multiplicative), and every combination of (E, T, S) specification is a distinct model in the family.

Implement SES by hand

Notice how the orange line (alpha=0.8) hugs every bump in the data, while the blue line (alpha=0.1) glides smoothly past most of the noise. Neither is universally better — the right alpha depends on how much of the observed variability is signal versus noise.

Fitting Holt-Winters with statsmodels

For production use you should let statsmodels fit the smoothing parameters automatically by minimizing the sum of squared one-step forecast errors.

from statsmodels.tsa.holtwinters import ExponentialSmoothing
import pandas as pd

# Assume `series` is a pandas Series with a DatetimeIndex and freq="MS"
model = ExponentialSmoothing(
    series,
    trend="add",            # additive trend (use "mul" for multiplicative)
    seasonal="add",         # additive seasonality
    seasonal_periods=12,    # 12 months per year
)
fit = model.fit()

forecast = fit.forecast(steps=12)   # 12-month horizon
print(fit.summary())

When ETS beats ARIMA

Both ARIMA and ETS are serious, widely-used forecasters. The practical rule of thumb:

ETS tends to shine when the seasonal pattern is stable and predictable — monthly retail, yearly energy demand, weekly call-center volume. ARIMA tends to win when the dynamics are driven by lagged shocks rather than a smooth underlying level-plus-trend-plus-season structure. In practice, the best strategy is to fit both and compare out-of-sample error on a held-out validation window.