Awesome

Conformal Time Series Decomposition

This repository contains the code and instructions for reproducing the experiments in the paper "Conformal time series decomposition with component-wise exchangeability".

Table of Contents

• Introduction
• Installation
• Usage
• Cite

Introduction

In this paper, we propose a novel method for decomposing time series data using conformal prediction and component-wise exchangeability. This README provides instructions on how to run the code and reproduce the experiments described in the paper.

Installation

To run the code, you need to have the following version of Python installed:

• Python 3.10.13

You can install the required dependencies by running the following command:

pip install -r requirements.txt

Usage

The runner.py script can be executed on any dataset using the --dataset argument. The dataset is decomposed into trend, seasonal, and noise components. Predictions are made on each of these components, including the original time series, using a specified regressor (--basemodel). Prediction intervals are calculated for each component. You can specify the conformal method for each component using the following arguments:

• --cp_method_original
• --cp_method_trend
• --cp_method_seasonal
• --cp_method_noise

Results are automatically logged to Weights & Biases (wandb) for each component, as well as for the original and recomposed time series.

When specifying each basemodel or conformal method, you may need to set hyperparameters. See the arg_parser in runner.py for more details.

Examples

Some examples with which some of our experimental results can be reproduced.

Synthetic Dataset

Using the Linear model, running Enbpi on the trend component, BinaryPoint on the seasonal component, and CV+ on the noise component:

python src/runner.py --dataset synthetic --basemodel Linear --cp_method_trend enbpi --cp_method_seasonal local_cp --cp_method_noise cv_plus --local_cp_method periodic --use_region False --use_exponential False

Energy Dataset

Using the MLP model, running ACI on the trend component, BinaryLocal on the seasonal component, and CV+ on the noise component:

python src/runner.py --dataset energy-consumption --basemodel MLP --cp_method_trend aci --cp_method_seasonal local_cp --cp_method_noise cv_plus --local_cp_method periodic --use_region True --use_exponential False

Sales Dataset

Using the Linear model, running ACI on the trend component, ExpLocal on the seasonal component, and CV+ on the noise component:

python src/runner.py --dataset sales --basemodel Linear --cp_method_trend aci --cp_method_seasonal local_cp --cp_method_noise cv_plus --local_cp_method periodic --use_region True --use_exponential True

Cite

If you find this work helpful, please cite

@misc{prinzhorn2024conformaltimeseriesdecomposition,
      title={Conformal time series decomposition with component-wise exchangeability}, 
      author={Derck W. E. Prinzhorn and Thijmen Nijdam and Putri A. van der Linden and Alexander Timans},
      year={2024},
      eprint={2406.16766},
      archivePrefix={arXiv},
      primaryClass={stat.ML},
      url={https://arxiv.org/abs/2406.16766}, 
}

Keywords

Time Series, Time Series Decomposition, Uncertainty, Conformal Prediction, Exchangeability Regimes, Machine Learning