gcastro-98/conformal-prediction

Conformal prediction applied to regression problems involving both exchangeable and time-series data. Carried out as part of (UB 2023) MSc thesis work.

Conformal prediction

Uncertainty quantification for distribution-free and data-agnostic problems is applied in terms of conformal prediction (CP) methodologies. In particular, this is applied to regression problems involving both exchangeable and time-series data. This work was carried out as part of (UB 2023) MSc thesis development.

Exchangeable data

Toy problem

Before dealing with a more complex datasets, we will study how CP performs in the toy problem proposed in this Kaggle discussion.

conformal-prediction's Issues

New visualization: statistical coverage vs. alpha value

A K-fold CV (with K=5 or 10) should be implemented to obtain K=5 or 10 values of the regression_coverage_score_v2 for each value of $\alpha$ miscoverage level. This way, the performance of each strategy can be studied in function of $\alpha$.

Implementing CQR both to the toy problem and the real regression problem

The CQR strategy should be implemented and compared to the rest of them (SCP, CV+, J+aB).

Implementing 'EnbPI' & visualizations for time-series problems

This notebook could be implemented and some former visualizations of ours, adapted to such purpose.

Adapting the toy problem to a real example

For instance, this notebook's dataset could be used (on which the tutorial for CQR regression was based, actually). They leverage the California Housing dataset from sklearn.