IC-538Chronos-T5's discrete probability prediction approach yields very small power-law exponents on NLL, limiting its scalability, and its in-distribution gains do not extend to out-of-distribution data
Qingren Yao, Chao-Han Huck Yang, Renhe Jiang, Yuxuan Liang, Ming Jin, Shirui Pan
The paper retrains Chronos-T5 and a simpler decoder-only transformer on the same 16.8B time-point corpus. Chronos-T5, which tokenises time series values into a discrete vocabulary and predicts the next token, exhibits power laws with very small exponents when evaluated with NLL, indicating limited scalability. The authors attribute this to the discrete probability prediction, where NLL on a discrete distribution is not distance-sensitive. On in-distribution data, Chronos-T5 shows a slight advantage in performance and scalability, but on out-of-distribution data the simpler decoder-only baseline performs better, and both models exhibit similar scaling patterns.
Evidence
correlational
Caveat
Chronos-T5 was retrained on the authors' 16.8B time-point corpus rather than its original training data. The authors note that NLL is not the most appropriate metric for evaluating discrete token prediction and that SMAPE is more suitable; specific exponent values are reported only in Figure 7, not in the text. Evaluation is restricted to univariate forecasting.