Daily Frequency and SHAP for Tiny Time Mixers

[fferrador]

Hi to you all,

I'm on the process of writing my master thesis and I would like to use Tiny Time Mixers on a proprietary dataset. However, this dataset only has daily frequency but all the examples I have seen use hourly frequency.

• How can I deal with this issue?
• Is TTM capable of making predictions with daily frequency? How can I do IT?

Additionally, I haven't seen any example of using SHAP to explain the output of TTM. Can it be done?

Thanks for your help.

[wgifford]

Hi @fferrador thanks for your interest in TSFM!

There is nothing preventing the use of TTM models on daily data. However, given the pre-trained models we have currently released the performance may be not be best. Most of the pretraining datasets we used for those models are at finer resolutions (hourly, minutely, etc.). In addition daily resolution with long context requirements may pose additional challenges as one or more years of historical data may not be available. The team is currently working on resolving some of these issues.

It is definitely possible to apply SHAP to time series forecasts; the team has done some work in this area: see, for example: https://arxiv.org/abs/2303.12316

If you have examples of applying explainability techniques to the output of TTM, it would be great if you would consider contributing them to our time series cookbook repo: https://github.com/ibm-granite-community/granite-timeseries-cookbook/tree/main