Conformal Predictions under Markovian Data
Authors: Frédéric Zheng, Alexandre Proutiere
ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our theoretical results are confirmed using numerical experiments, both on synthetic and real-world data (e.g., for the prediction on the EUR/SEK exchange rate). |
| Researcher Affiliation | Academia | 1Division of Decision and Control Systems, EECS KTH and Digital Futures, Stockholm, Sweden. |
| Pseudocode | No | The paper describes the methods narratively and mathematically, but does not include any structured pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not contain any explicit statements about the release of source code for the described methodology, nor does it provide a link to a code repository. |
| Open Datasets | Yes | The dataset can be found at https://www.histdata. com/ (referring to EUR/SEK exchange rate). In this second example, we consider the same dataset as (Zaffran et al., 2022), which contains the French electricity price between 2016 to 2019, reported every hour. |
| Dataset Splits | Yes | At each timestep, we apply conformal prediction to the next return rt+1 with a rolling window of fixed size divided into training and calibration datasets (1 month = 30x24x60 data points for each in this example). In this second example, we consider the same dataset as (Zaffran et al., 2022), which contains the French electricity price between 2016 to 2019, reported every hour. ... with a rolling window of fixed size (18 months = 18x30x24 data points for both training/calibration). |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., CPU, GPU models, or memory) used for running the experiments. |
| Software Dependencies | No | The paper does not provide specific version numbers for software dependencies or libraries used in the experiments. |
| Experiment Setup | Yes | An experiment consists in generating one trajectory of length N +n+1 and in applying CP to the last point. We repeat the experiment Ntrials = 1000 times and report the average coverage rate. We fix N = 10000. For a given true model µ(x) and independent symmetric noise εt, we generate Yt = µ(Xt) + εt (refer to Appendix B.1 for details). The classical AR(1) models are reversible Markov chains defined by the following recursive equation: n, Xn+1 = θXn + εn+1 with εn N(0, ω2) and for some θ [0, 1[ and ω > 0. For example, θ = 0.9, ω = 1 for Gaussian AR. For real-world data, they use a rolling window of fixed size divided into training and calibration datasets (1 month = 30x24x60 data points for each in this example) or (18 months = 18x30x24 data points for both training/calibration). |