Which model for forecasting with seasonality and trend?

What's really being asked

This question tests your ability to choose an appropriate, simple model based on data characteristics before reaching for more complex solutions. It's a check for foundational knowledge and a structured approach to modeling. The interviewer wants to see if you can identify the Holt-Winters method as a powerful baseline and explain how its components directly map to the trend and seasonality described in the problem.

The full answer

A strong answer identifies the Holt-Winters seasonal method and explains its three core components. First, the level (ℓt), which is the smoothed value of the series. Second, the trend (bt), which captures the rate of change, addressing the upward trend. Third, the seasonal component (st), which captures the repeating weekly pattern. A candidate should also mention the smoothing parameters alpha, beta, and gamma. Finally, they must explain the choice between the additive method (for constant seasonal variations, e.g., sales increase by 100 units every Monday) and the multiplicative method (for variations proportional to the level, e.g., sales increase by 10% every Monday).

The mistakes people make

The biggest red flag is immediately suggesting a complex model like Prophet, a neural network, or even SARIMA without first mentioning Holt-Winters. The question specifically asks for a classical baseline. Naming Holt-Winters but being unable to explain how the level, trend, and seasonal components work is another failure pattern; it suggests memorization without understanding. Finally, being unable to articulate the practical difference between additive and multiplicative seasonality shows a lack of depth.

What usually comes next

Expect follow-ups like: "How would you choose between the additive and multiplicative models in practice?" (Answer: Visualize the data; if seasonal variance grows with the level, use multiplicative. Otherwise, compare model fit on a validation set using AIC/BIC). Another likely question is: "For daily data with a weekly pattern, what is the seasonal period 'm'?" (Answer: m=7). They might also ask about the model's limitations, such as handling multiple seasonalities (e.g., weekly and yearly) or incorporating external holiday effects.

A concrete example

For daily inventory data with a weekly pattern, the Holt-Winters model would use a seasonal period of m=7. The model would learn a baseline level, an upward trend (e.g., demand growing by 10 units per day), and a set of 7 seasonal factors. If using a multiplicative model, the forecast for a Monday might be (level + trend) 1.2, while a Saturday might be (level + trend) 0.7, reflecting the consistent weekly pattern on top of the overall growth.