The question "which model is best for demand forecasting" has no stable answer. It depends on how much history you have, how many series you forecast, and how much of the signal lives outside the series itself.
Three families dominate practice, and each has a regime where it is hard to beat.
Statistical models
Exponential smoothing, ARIMA and their seasonal variants model a single series in terms of level, trend and seasonality.
They win when history is short, the series is reasonably stable, and there is little external information to exploit. They are fast, they need almost no tuning, and they are difficult to beat on smooth, high-volume series.
They struggle when demand depends on drivers outside the series, such as price or promotional calendars, and when the series is intermittent.
Gradient boosting
Tree-based models such as gradient boosting treat forecasting as a supervised learning problem: build features from the history and from external drivers, then predict the next periods.
They win when you have many series and useful covariates. Promotions, prices, holidays, weather and product attributes all enter naturally as features, and a single model can learn patterns shared across thousands of items.
They need careful feature engineering, particularly around lags, and they do not extrapolate trends beyond the range they saw in training.
Deep learning
Neural architectures for time series learn shared structure across a large panel of series and can capture longer-range and non-linear dependencies.
They win at scale: many series, long history, complex seasonality and interactions. They also handle cold-start items reasonably well, because the pattern is learned across the panel rather than from one item's history.
They need more data and more compute, and they are the least forgiving family when the input data is inconsistent.
Why combining beats choosing
The families fail in different places. Statistical models are stable but blind to drivers. Boosted trees exploit drivers but flatten trends. Neural models capture complex structure but overfit thin history.
Evaluating candidates per series with proper backtests, then combining the ones that hold up, is more robust than committing to a single family across an entire catalogue. The combination is rarely the best model on any individual series, and it is usually the best portfolio across all of them.
What this means in practice
You should not have to pick. What matters is that the evaluation is honest: out-of-sample, on your own history, compared against a naive baseline, and reported per series rather than as one aggregate number that hides where the model fails.