Exponential Smoothing & ETS
SES, Holt, Holt-Winters, ETS state space, connection to ARIMA, MLE tuning
ARIMA requires ACF/PACF identification to choose p, d, q — a manual process that breaks at scale and fails for practitioners without time series expertise. Exponential smoothing methods sidestep this by imposing a simple structural assumption: past observations should be weighted by recency, with exponentially decaying weights. SES does this for level; Holt's method adds a trend component; Holt-Winters adds seasonality.
The mistake is treating these as heuristic update rules. ETS (Error-Trend-Seasonality) is a proper statistical state space model — it has a likelihood function, parameters estimated via MLE, and AIC-based model selection across 30 component combinations.
This means ETS implicitly performs ARIMA order selection without the ACF/PACF identification step, which is why ETS consistently outperforms ARIMA on large benchmark datasets like M3 and M4.
Key points
- ARIMA order selection requires ACF/PACF interpretation, which is manual, error-prone, and impossible at scale. ETS bypasses this entirely: it assumes exponentially decaying weights on past observations, estimates the decay parameter α via MLE, and selects among Error-Trend-Seasonality component combinations via AIC. The automation is the point — ETS delivers ARIMA-quality forecasts without per-series manual identification.
- Simple Exponential Smoothing (SES): ŷ_{t+1} = αY_t + (1-α)ŷ_t = Σ α(1-α)^j Y_{t-j}. α ∈ (0,1) controls recency weighting — α=1 is a naïve forecast (only yesterday matters), α→0 weights all history equally. Optimal α is estimated by MLE. SES is the optimal forecast for a random walk with Gaussian noise — it is exactly equivalent to ARIMA(0,1,1) with θ = α-1. A high fitted α (e.g., 0.92) means the series has near-random-walk dynamics: past values beyond one period ago carry almost no predictive information.
- Holt's double exponential smoothing adds a trend component. Level: L_t = αY_t + (1-α)(L_{t-1} + b_{t-1}). Slope: b_t = β(L_t - L_{t-1}) + (1-β)b_{t-1}. Forecast: ŷ_{t+h} = L_t + hb_t. Linear trend extrapolation at long horizons is dangerous — a series that has been trending up for 2 years is not guaranteed to continue. The damped trend variant (damped-trend Holt) multiplies the slope by φ^h with φ ∈ (0.8, 0.98): the slope shrinks toward zero at long horizons. This is one of the most reliable improvements for medium-horizon forecasting.
- Holt-Winters adds a seasonal component. Additive: S_t = γ(Y_t - L_{t-1} - b_{t-1}) + (1-γ)S_{t-s}. Multiplicative: S_t = γ(Y_t / (L_{t-1} + b_{t-1})) + (1-γ)S_{t-s}. All parameters (α, β, γ) are estimated jointly by MLE — not independently. Multiplicative seasonality handles the case where seasonal amplitude scales with level; log-transforming and using additive is an equivalent approach.
- ETS(E, T, S) notation: E ∈ {A, M} for error type (additive or multiplicative), T ∈ {N, A, Ad, M, Md} for trend (none, additive, additive damped, multiplicative, multiplicative damped), S ∈ {N, A, M} for seasonality. Up to 30 valid combinations. AIC over all valid ETS variants selects the best automatically — this is the equivalent of Box-Jenkins order selection done via likelihood comparison rather than ACF/PACF reading.
- ETS-ARIMA equivalence proves that exponential smoothing is not ad hoc. SES (ETS(A,N,N)) = ARIMA(0,1,1). Holt's method (ETS(A,A,N)) = ARIMA(0,2,2). Holt-Winters additive (ETS(A,A,A)) = ARIMA(0,1,m+1)(0,1,0)[m]. ETS implicitly selects ARIMA order via structural assumptions rather than ACF/PACF exploration. This equivalence explains why ETS and ARIMA perform similarly on benchmark datasets — they're fitting the same class of models via different parameterisations.
- ETS outperforms ARIMA in practice for two reasons. Parsimony: ETS has fewer parameters with stronger regularisation via structural assumptions, reducing overfitting on short series. Multiplicative error models: ETS(M,...) handles heteroskedasticity (variance increasing with level) naturally by multiplying the noise term by the current level. ARIMA with Gaussian errors assumes constant variance, which misspecifies series with growing amplitude. Holt-Winters multiplicative error produces correct prediction intervals for heteroskedastic series where ARIMA does not.
- Croston's method handles intermittent demand (series with many zeros) — separately exponentially smooths demand size and inter-demand interval. SES and Holt-Winters produce nonzero forecasts even when zeros dominate, which is wrong by construction for spare-parts or SKU-level intermittent series. For series with more than 30-50% zeros, Croston or ADIDA (Aggregated-Disaggregated Intermittent Demand Approach) are the correct starting points.
ETS is not a heuristic — it is a proper state space model whose parameters are estimated by MLE and whose AIC-selected variant implicitly performs ARIMA order selection without the ACF/PACF identification step. The ETS-ARIMA equivalence (SES = ARIMA(0,1,1), Holt = ARIMA(0,2,2)) proves these are the same underlying model class in different parameterisations. The practical M4 conclusion: always ensemble ETS with ARIMA and a simple baseline, because model uncertainty across forecast horizons is large enough that combination consistently dominates any single method.
Recap
- Exponential smoothing = recency-weighted average, sidesteps ACF/PACF identification.
- SES (level) → Holt (+trend) → Holt-Winters (+seasonality); all params estimated jointly by MLE.
- Damped trend (slope × φ^h, φ∈0.8–0.98) — most reliable medium-horizon improvement.
- ETS(E,T,S) = proper state space model; AIC over ~30 variants auto-selects (Box-Jenkins via likelihood).
- ETS-ARIMA equivalence: SES=ARIMA(0,1,1), Holt=ARIMA(0,2,2) — same model class, different parameterisation.
- Multiplicative-error ETS handles heteroskedasticity (variance grows with level) where Gaussian ARIMA misspecifies.
- Croston's for intermittent demand (>30-50% zeros); M4 lesson: always ensemble ETS+ARIMA+baseline.
Check your understanding
Q1. You fit SES to a daily sales series and find the optimal α = 0.92 via MLE. What does this imply about the data and what model class does SES correspond to at this parameter value?
- A) α ≈ 0.92 means almost all weight is on recent data, implying near-random-walk dynamics. SES with α ≈ 1 corresponds to ARIMA(0,1,1) with θ = α−1 ≈ −0.08 — forecasts collapse to the last observation.
- B) α = 0.92 is an unusually high value indicating the MLE optimisation has converged to a poor local minimum; constrain α to the range [0.1, 0.5] and re-fit to obtain a more reliable parameter estimate.
- C) α = 0.92 indicates strong long-range autocorrelation, because the high weight placed on recent observations causes information from many past periods to gradually accumulate in the smoothed level.
- D) SES with α = 0.92 corresponds exactly to ARIMA(1,0,0) with φ = 0.92, because exponential smoothing and autoregression are mathematically equivalent constructions at all values of α.
Q2. ETS(M,A,M) has multiplicative error, additive trend, multiplicative seasonality. Which TWO statements about it versus ETS(A,A,A) are correct?
- A) ETS(M,A,M) has multiplicative, heteroskedastic errors implying a log-normal-like forecast distribution whose variance grows with the level, unlike ETS(A,A,A)'s constant-variance Gaussian errors.
- B) Because ETS(M,A,M)'s error variance is not constant, its prediction intervals must generally be computed via simulation rather than the closed-form analytical Gaussian formula that applies to ETS(A,A,A).
- C) ETS(M,A,M) uses additive errors internally despite the M notation; the multiplicative label refers only to seasonality, so its interval formulas are identical to ETS(A,A,A)'s in every respect.
- D) ETS(M,A,M) produces systematically narrower intervals than ETS(A,A,A), because multiplicative errors self-correct — deviations from trend become proportionally smaller as the overall level grows.
Q3. The M4 forecasting competition showed that ETS and ARIMA individually perform worse than simple combination methods. How does this affect how you should use ETS in production?
- A) Model uncertainty is large enough that ensembling reduces variance without proportionally increasing bias. In production: ensemble ARIMA+ETS+Prophet, add a theta baseline, weight equally or via validation.
- B) The M4 combination result applies only to the specific dataset distribution used inside that competition; for domain-specific series like retail or energy, a single well-tuned ETS model always outperforms naive combinations.
- C) The M4 result means ETS should be abandoned entirely in favour of gradient-boosted tree models; the competition proved statistical methods are outclassed by machine learning regardless of combination strategy used.
- D) Use ETS only ever as a component within combinations, never as a standalone model; always combine with at least 5 other distinct methods to reproduce the M4 accuracy gains reliably in production.
Try it interactively
ML Systems Lab is a free interview-prep platform for ML engineers — work through the full interactive module, quizzes, and drills.
Open ML Systems Lab →