Adapting language model architectures for time series forecasting

Tokenizing time series data and treating it like a language enables a model whose zero-shot performance matches or exceeds that of purpose-built models.