What is the difference between classical statistical time series models and deep learning models?

Classical models assume linear structure and stationarity, while deep learning models learn nonlinear, high-dimensional temporal patterns automatically.

Classical models like ARIMA, SARIMA, and VAR rely on explicit assumptions such as stationarity, linearity, and predefined lag structure. They require manual feature engineering (lags, differencing, seasonality removal). Deep learning models like LSTM, GRU, and Temporal Convolutional Networks learn representations directly from raw sequences, capturing nonlinear dependencies and long-range interactions. However, they require more data and are less interpretable.