What neural approaches add, and what they cost
Deep learning models built for sequential data, and more recently transformer-based and LLM-adjacent forecasting approaches, can in principle capture patterns that classical statistical models can't: complex nonlinear relationships, long-range dependencies across many past time steps, and the ability to learn jointly across many related series at once (forecasting demand for thousands of products together, sharing statistical strength across them, rather than fitting a separate ARIMA model to each). That last capability, learning across many related series simultaneously, is often the strongest practical argument for a neural approach, since classical models fit one series at a time and can't borrow patterns from similar series the way a shared neural model can.
That capacity comes at a real cost. Neural models have far more parameters to fit, which means they need substantially more data to avoid simply memorizing noise in the training set rather than learning genuine structure. They're also more expensive to train and tune, and far harder to interpret than an ARIMA model whose coefficients you can inspect directly. None of this makes neural approaches wrong, it makes them a tool with a narrower set of conditions under which they're the right one.