Always start with a simple baseline (last period, seasonal naive) before trying complex models. If you cannot beat the baseline, the complex model is not adding value.
Strong answers cover: decomposing time series into trend, seasonality, and residuals, choosing appropriate models (moving averages, exponential smoothing, ARIMA, Prophet), handling stationarity, cross-validation with time-based splits (not random), and accuracy metrics (MAE, MAPE, RMSE). Best candidates discuss when simple methods outperform complex ones and how to communicate forecast uncertainty.
Advanced analytical skill. Tests statistical depth and practical judgment. Ask: "How do you communicate forecast uncertainty to stakeholders who want a single number?" to gauge communication maturity.