Time series forecasting has seen major advances with neural and foundation model approaches. Classical baselines (always compare to these): ARIMA/SARIMA for univariate, Exponential Smoothing (ETS) for seasonal data, Prophet for business metrics with seasonality and holidays. Neural approaches: 1) N-BEATS: interpretable neural basis expansion with trend and seasonality decomposition stacks. Excels on M4 competition data, good for business metrics. 2) PatchTST: transformer-based, treats time series patches as tokens (like images), achieves strong performance on ETT datasets. 3) DLinear: embarrassingly simple linear model that often beats complex transformers on standard benchmarks. 4) TimesNet: converts 1D time series to 2D space to leverage 2D convolutional networks. Foundation models (zero-shot forecasting): Chronos (Amazon): transformer trained on diverse time series data. Works zero-shot on new domains. Lag-Llama: causal language model for probabilistic forecasting. TimeGPT: first foundation model for time series, competitive zero-shot performance. Model selection: for univariate with strong seasonality use Prophet or ETS. For multivariate with long history use neural models. For cold-start (limited history) use foundation models. Evaluation: MAPE, RMSE, MASE, wQL (weighted quantile loss for probabilistic forecasts). Always evaluate on held-out test set respecting temporal ordering.