Adaptive Conformal Predictions for Time Series

Authors: Margaux Zaffran, Olivier Feron, Yannig Goude, Julie Josse, Aymeric Dieuleveut

ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We lead extensive fair simulations against competing methods that advocate for ACI s use in time series. We conduct a real case study: electricity price forecasting.
Researcher Affiliation Collaboration 1Electricit e de France R&D, Palaiseau, France 2INRIA Sophia Antipolis, Montpellier, France 3CMAP, Ecole Polytechnique, Institut Polytechnique de Paris, Palaiseau, France 4Fi ME, Universit e Paris-Dauphine, France 5LMO, Universit e Paris-Saclay, Orsay, France 6IDESP, Montpellier, France.
Pseudocode Yes Algorithm 1 Online Expert Aggregation on ACI (Ag ACI)
Open Source Code Yes All the code and data to reproduce the experiments are made available on Git Hub.
Open Datasets No The data set contains the French electricity spot prices, set by an auction market, from 2016 to 2019.
Dataset Splits Yes We predict for the year 2019, using a sliding window of 3 years, as described in Figure 5(a), using one year and 6 months as proper training set and the most recent year and a half for calibration.
Hardware Specification No The paper does not mention any specific hardware details such as CPU, GPU models, or memory specifications used for the experiments.
Software Dependencies No To allow for better benchmarking of existing and new methods, we provide (re-)implementations in Python of (all) the described methods and a complete pipeline of analysis on Git Hub. As explained in Section 4, the code for Ag ACI is, for now, the only one available only in R.
Experiment Setup Yes All the random forests model have the same parameters, that are the following: Number of trees: 1000 Minimum sample per leaf: 1 (default) Maximum number of features: d (default)