Chapter 8 (Exponential smoothing) Flashcards
Simple Exponential Smoothing (SES)
forecasting data with no clear trend or seasonal pattern
ETS(A,N,N)
- Error: Additive
- Trend: None
- Season: None
Holt’s linear trend method
Allow forecasting of data with a trend
ETS(A,A,N)
- Error: Additive
- Trend: Additive
- Season: None
Damped trend method
Forecasting of data that ‘dampens’ the trend so that it approaches a constant some time in the future
0.8 <= Damp value <= 0.98
ETS(A,Ad,N)
- Error: Additive
- Trend: Additive damped
- Season: None
ETS methods with seasonality
Additive vs multiplicative
- Additive method: is preferred when the seasonal variations are roughly constant through the series
- Multiplicative method: is preferred when the seasonal variations are changing proportional to the level of the series
Holt-Winters’ additive method
ETS(A,A,A)
- Error: Additive
- Trend: Additive
- Season: Additive
Holt-Winters’ multiplicative method
ETS(M,A,M)
- Error: Multiplicative
- Trend: Additive
- Season: Multiplicative
If season is multiplicative, error must also be multiplicative.
Having additive error with multiplicative season may result to numeric instability
Holt-Winters’ damped method
ETS(M,Ad,M)
- Error: Multiplicative
- Trend: Additive damped
- Season: Multiplicative
ETS models - general notation
- Error: Additive (A) or Multiplicative (M)
- Trend: None (N), Additive (A) or Additive damped (Ad)
- Seasonality: None (N), Additive (A) or Multiplicative (M)
Model selection
AIC determines which ETS model is most appropriate
AIC only works with same model types
Prediction intervals between additive and multiplicative methods
Even if same smoothing parameters, the point forecasts will be the same but prediction intervals will differ
Smoothing parameters
- Alpha (α): if close to 1, then level changes rapidly
- Beta (β): if close to 1, then trend changes often
- Gamma (γ): if close to 1, then seasonality updates faster
- Phi (Φ): damping, often between 0.8 to 0.98