How does next-token prediction work for time series data?