Skip to content
Tutorials Center
My learning

Topics

Lesson 19 of 40 ยท Smoothing and Exponential Methods

ETS State-Space Models

Understand the ETS (Error, Trend, Seasonal) taxonomy of exponential smoothing, fit ETSModel in statsmodels, select models with AIC, and get proper prediction intervals.

The methods in this chapter (SES, Holt, Holt-Winters) were invented as practical recipes. Later, statisticians showed that each one is the point forecast of a statistical model with an explicit error term. Those models are called ETS models, for Error, Trend, Seasonal. The statistical view gives you things the recipes could not: likelihood-based parameter estimation, information criteria for comparing models, and prediction intervals.

The ETS naming scheme

An ETS model is written ETS(E, T, S), where each letter takes one of:

Component Options
Error A additive, M multiplicative
Trend N none, A additive, Ad additive damped, M multiplicative, Md damped multiplicative
Seasonal N none, A additive, M multiplicative

So ETS(A,N,N) is simple exponential smoothing with additive errors, ETS(A,A,N) is Holt's linear method, and ETS(A,Ad,A) is damped Holt-Winters with additive seasonality. With 2 error types, up to 5 trend types and 3 seasonal types there are up to 30 combinations (some are unstable and excluded).

ETS = Error, Trend, Seasonal

Error Additive or multiplicative.
Trend None, additive, damped, multiplicative.
Season None, additive, multiplicative.

ETS(A,N,N) is simple exponential smoothing.

1 / 2

What the error type changes

Point forecasts of ETS(A,...) and ETS(M,...) can be identical for the same smoothing parameters, but the error term differs: additive errors have constant size, multiplicative errors are a percentage of the level. That affects the parameter estimates, the likelihood and, most visibly, the width of the prediction intervals, which grow with the level in the multiplicative case. Multiplicative errors are natural for positive series whose noise scales with the level.

State-space form

Every ETS model has the form: observation = function of states + error, new states = function of old states + (smoothing parameter) times error. For ETS(A,N,N):

  • Observation: y(t) = l(t-1) + e(t)
  • State: l(t) = l(t-1) + alpha * e(t)

This matches the error-correction form of SES from the first lesson of this chapter. The states (level, slope, seasonal values) are never observed directly; they are filtered out of the data.

Fitting with ETSModel

ETSModel in statsmodels.tsa.exponential_smoothing.ets uses the full taxonomy and exposes the likelihood, AIC and prediction intervals.

Python Fit an ETS(A,Ad,A) model

Press Run (or Ctrl + Enter) to see the result here.

Program input ยท what you'd type while the program runs

Type the answers before you press Run, one per line. Each time the program reads a line (with input()), it gets the next line from this box.

summary_frame gives the mean forecast and the lower and upper bounds of a 95 percent prediction interval in one table, something Holt-Winters recipes could not offer directly.

Selecting a model with AICc

Because ETS models are fitted by maximum likelihood, you can compare them with an information criterion: AIC, or AICc (corrected for small samples), where lower is better. Fit several candidates and choose the smallest AICc, then confirm on held-out data.

Python Choosing among ETS candidates by AICc

Press Run (or Ctrl + Enter) to see the result here.

Program input ยท what you'd type while the program runs

Type the answers before you press Run, one per line. Each time the program reads a line (with input()), it gets the next line from this box.

The multiplicative variants win clearly here because the seasonal swings grow with the level and the noise is proportional. Always compare models fitted to the same data and same transformation: AIC values from log-transformed and untransformed data are not comparable.

Automatic selection

R's ets() and Python packages such as statsforecast (AutoETS) loop over the allowed combinations and pick the best AICc automatically. With statsmodels you write that loop yourself, as above. Automatic selection works well for large collections of series; for one important series, still look at the plot of forecasts and residuals.

Residual diagnostics

A fitted model is only useful if its residuals look like white noise: no autocorrelation, roughly constant variance. Check with the ACF and the Ljung-Box test.

Python Residual check for an ETS model

Press Run (or Ctrl + Enter) to see the result here.

Program input ยท what you'd type while the program runs

Type the answers before you press Run, one per line. Each time the program reads a line (with input()), it gets the next line from this box.

ETS or ARIMA?

Both families are workhorses. ETS models describe trend and seasonality directly in terms of level, slope and seasonal states, and are easy to explain. ARIMA models (next chapter) describe the autocorrelation structure and are better when the data show rich short-term dynamics. They overlap (SES equals ARIMA(0,1,1)) but each also contains models the other lacks, so on important series it is worth fitting both and comparing on held-out data.

Check your understanding

Quick check ยท 4 questions

of 4 answered You scored / 4 ยท perfect!

  1. 1. What does ETS(A,Ad,N) stand for?

    A is additive error, Ad is damped additive trend and N means no seasonal component: damped Holt.
  2. 2. What do ETS models offer that the original Holt-Winters recipes did not?

    The state-space formulation attaches a probability model to the method.
  3. 3. When is a model with multiplicative errors natural?

    Multiplicative errors scale with the level, so interval widths grow as the series grows.
  4. 4. Which comparison is invalid?

    The likelihoods are on different scales when the data are transformed, so AIC values are not comparable.

Summary

  • ETS names a model by Error, Trend and Seasonal types; SES, Holt and Holt-Winters are special cases.
  • State-space form gives maximum likelihood estimation, AIC/AICc selection and proper prediction intervals.
  • ETSModel in statsmodels supports additive/multiplicative errors, damped trends and seasonality.
  • Compare candidates by AICc on the same data and confirm with held-out accuracy.
  • Check residuals for leftover autocorrelation, then move on to ARIMA for richer dynamics.

Next: Autoregressive (AR) Models