Climate Informatics 2023: 'A locally time-invariant metric for climate model ensemble predictions of extreme risk'

Code and data associated with the submission 'A locally time-invariant metric for climate model ensemble predictions of extreme risk'.

Setup: clone this repository and use conda env create -f conda_environment.yml to install relevant dependencies.

1. Data download

Data used to produce the results presented in tha paper are provided in the folder 'data'. These are time-series taken from one realisation per model for the CMIP6 members GFDL-ESM4, IPSL-CM6A-LR, MPI-ESM1-2-HR, MRI-ESM2-0, and UKESM1-0-LL, for nine cities Paris, Chicago, Sydney, Tokyo, Kolkata, Kinshasa, Shenzhen and Santo Domingo. The same time-series from an observational reanalysis data, W5E5, are also provided.

Analysis for other spatial locations, model realisations, climate variables, etc. can be run by downloading the relevant data using the notebook data_download.ipynb

2. Locally time-invariant permutations

A simple demonstration of the locally time-invariant permutation approach to model evaluation described in the paper is provided in metric_demo.ipynb.

3. Results

The Bayesian model averaging weights and evaluation presented in the paper can be reproduced using the notebook bayesian_model_averaging.ipynb.