Sequential Predictive Conformal Inference (SCPI) for Time Series

[Kevin-Chen0]

Describe the feature request you'd like Include SCPI as a second, alternative method within the MapieTimeSeriesRegressor class alongside the existing EnbPI method.

Why is this beneficial? SCPI improves upon EnbPI by including serial dependence across prediction residuals (or conformity score), aka sequential conformal prediction. This can further reduce the prediction interval width for the same data and empirical coverage compared to EnbPI.

Additional context Xu, C. and Xie, Y. Sequential predictive conformal inference for time series. arXiv preprint arXiv:2212.03463, 2022b. [link]

[LacombeLouis]

Hi @Kevin-Chen0, Thank you for your comment. Indeed, I think this could be a valuable addition to MAPIE. For the moment, we have focused our efforts in Time Series on the following papers (#334):

• Gibbs, I., & Candes, E. (2021). Adaptive conformal inference under distribution shift. Advances in Neural Information Processing Systems, 34, 1660-1672.
• Zaffran, M., Féron, O., Goude, Y., Josse, J., & Dieuleveut, A. (2022, June). Adaptive conformal predictions for time series. In International Conference on Machine Learning (pp. 25834-25866). PMLR.

If you want to contribute to the library, I think this could be a very interesting new addition! Don't hesitate to contact us if you have any questions regarding the how and where to implement this method.

[Kevin-Chen0]

Hi @LacombeLouis,

Yes, I can contribute to MAPIE by adding SCPI there. Will you also be including ACI in MapieTimeSeriesRegressor? That way, we can ensure I follow the same structure when incorporating SCPI.

Thanks! ~Kevin

[LacombeLouis]

Hi @Kevin-Chen0,

Indeed, we are including it at this very moment, it's in the final stages of development... 🔜 https://github.com/scikit-learn-contrib/MAPIE/pull/341.

I think that as soon as it's released, we will make sure to keep you posted! Thank you! Louis

[valeman]

Hi @Kevin-Chen0, Thank you for your comment. Indeed, I think this could be a valuable addition to MAPIE. For the moment, we have focused our efforts in Time Series on the following papers (#334):

• Gibbs, I., & Candes, E. (2021). Adaptive conformal inference under distribution shift. Advances in Neural Information Processing Systems, 34, 1660-1672.
• Zaffran, M., Féron, O., Goude, Y., Josse, J., & Dieuleveut, A. (2022, June). Adaptive conformal predictions for time series. In International Conference on Machine Learning (pp. 25834-25866). PMLR.

If you want to contribute to the library, I think this could be a very interesting new addition! Don't hesitate to contact us if you have any questions regarding the how and where to implement this method.

I think SPCI outperforms these methods as who's in SPCI paper?