Nixtla's Open Source Time Series Ecosystem.

Nixtla Community

FYI: Quick singular use case using 2-years worth of data (1-year input and 1-year test). Compared a highly tweaked (and trained) Neural-Prophet model versus out of the box, no tweaks one-shot TimeGPT;  the MAPE difference was only 1.4% higher using TimeGPT.  Great job Nixtla.

That's amazing! Probably worth trying <https://nixtlaverse.nixtla.io/nixtla/docs/tutorials/finetuning.html|fine-tuning TimeGPT> a bit and see if it improves further! Let us know if you try that.

Thanks Marco. What is odd is that when holidays are included, the MAPE decreases from only 8.47 to 8.33. I had higher expectations. However, after thinking about it, given how TimeGPT is trained on many other time-series data, holidays clearly do not affect all time-series equally.

And fine tuning (10, 20 ...) raised the MAPE as well. My guess is that I am not applying fine tuning correctly.
Still quite amazed at TimeGPS's "out of the box" forecasting accuracy.

Screenshot 2024-12-16 at 17.36.36.png

Here are results using "out-of-box" TimeGPT settings. Not bad. Some holidays are significant anomalies in this business vertical

Hmm, I see! Maybe fine-tuning makes the model overfit. You could also try specifying `finetune_loss='mape'` along with `finetune_steps=10` and see if you can get more improvements.