Which classical baseline model handles weekly seasonality and upward trend?

WHAT THIS TESTS: This question tests whether you can map data characteristics to classical model architecture. The interviewer wants to see that you understand how exponential smoothing models decompose a signal into level, trend, and seasonal components, and that you know when to reach for Holt-Winters over simpler methods. It also checks if you understand the additive versus multiplicative distinction and whether you treat ETS as a baseline before jumping to more complex models.

A GOOD ANSWER COVERS: First, name Holt-Winters triple exponential smoothing as the natural baseline. Second, explain the three smoothing equations: the level equation updates the local mean with smoothing parameter alpha, the trend equation updates the slope with parameter beta star, and the seasonal equation updates the periodic deviation with parameter gamma. Third, specify that weekly seasonality means setting the seasonal period m to 7. Fourth, state the additive versus multiplicative choice: additive when seasonal swings are roughly constant in absolute units, multiplicative when they scale proportionally with the level of demand. Fifth, note that the forecast equation combines these three components to project forward.

COMMON WRONG ANSWERS: Jumping straight to SARIMA or machine learning without mentioning exponential smoothing is a red flag because it ignores the canonical baseline for structured trend plus seasonality. Confusing double exponential smoothing with triple exponential smoothing is another error; Holt's method handles trend but not seasonality, so it would miss the weekly cycle. Failing to mention the period m equals 7 suggests you are reciting formulas without connecting them to the problem. Claiming that classical models cannot handle trend and seasonality together reveals a gap in fundamentals.

LIKELY FOLLOW-UPS: The interviewer may ask how you would choose between additive and multiplicative seasonality in practice, so be ready to say you would inspect seasonal subseries plots or compare AICc across fitted models. They may ask how to handle multiple seasonalities, such as weekly plus monthly patterns, which classical Holt-Winters does not capture natively. They might also ask how you would set or tune the smoothing parameters, so mention maximum likelihood estimation or minimizing one-step forecast errors. Finally, expect a question on prediction intervals and why ETS generates them naturally from the state space formulation.

ONE CONCRETE EXAMPLE: Suppose you are forecasting daily widget sales. You fit an ETS model with additive seasonality and damped trend. The level component settles around one thousand units after smoothing out daily noise. The trend component estimates an increase of ten units per day. The seasonal component estimates that Mondays are two hundred units below the level and Saturdays are three hundred units above. Because you chose additive seasonality, those Monday and Saturday deviations stay fixed at roughly two hundred and three hundred units even as the baseline grows. If demand variability instead scaled with volume, you would switch to multiplicative seasonality so that Saturday peaks become roughly thirty percent above level rather than a flat three hundred units.

Source: otexts.com