Generative AI and foundation models are quickly becoming central in many areas of data science. However, their role in time series forecasting is still not fully clear. Are these models bringing real improvements, or are they just another layer of complexity? In this talk, I use time series forecasting as a concrete setting to explore this question. I compare traditional statistical models, recurrent neural networks, and more recent Transformer-based approaches that are often linked to the foundation model paradigm. The analysis includes traditional models such as ARIMA, ETS and Prophet, recurrent neural networks such as the LSTM and GRU, and transformer-based models such as Informer, Autoformer, and FEDformer. The comparison is based on benchmark datasets from energy, finance, and transportation. I look at both point forecasts and uncertainty, using bootstrap prediction intervals for the neural network models and comparing them with classical approaches. Rather than focusing only on performance metrics, the goal is to understand where these newer models add value, and where simpler methods remain competitive. I will also discuss some of the main challenges behind using generative AI in this context, including interpretability, computational cost, and reliability.
