Ecosyste.ms: Awesome
An open API service indexing awesome lists of open source software.
https://github.com/paulnorthrop/lite
Likelihood-Based Inference for Time Series Extremes
https://github.com/paulnorthrop/lite
clustered extremal-index extreme-value-statistics extremes frequentist generalised-pareto inference likelihood log-likelihood threshold time-series
Last synced: 26 days ago
JSON representation
Likelihood-Based Inference for Time Series Extremes
- Host: GitHub
- URL: https://github.com/paulnorthrop/lite
- Owner: paulnorthrop
- License: gpl-3.0
- Created: 2022-04-04T16:47:25.000Z (almost 3 years ago)
- Default Branch: main
- Last Pushed: 2024-07-17T20:16:02.000Z (6 months ago)
- Last Synced: 2024-07-19T01:09:34.709Z (6 months ago)
- Topics: clustered, extremal-index, extreme-value-statistics, extremes, frequentist, generalised-pareto, inference, likelihood, log-likelihood, threshold, time-series
- Language: R
- Homepage: https://paulnorthrop.github.io/lite/
- Size: 4.53 MB
- Stars: 3
- Watchers: 3
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.Rmd
- Changelog: NEWS.md
- License: LICENSE.md
Awesome Lists containing this project
README
---
output: github_document
---
```{r, echo = FALSE}
knitr::opts_chunk$set(
collapse = TRUE,
comment = "#>",
fig.path = "README-"
)
```
# lite
[](https://ci.appveyor.com/project/paulnorthrop/lite)
[](https://github.com/paulnorthrop/lite/actions/workflows/R-CMD-check.yaml)
[](https://app.codecov.io/github/paulnorthrop/lite?branch=main)
[](https://cran.r-project.org/package=lite)
[](https://cran.r-project.org/package=lite)
[](https://cran.r-project.org/package=lite)
## Likelihood-Based Inference for Time Series Extremes
The **lite** package performs likelihood-based inference for stationary time series extremes. The general approach follows [Fawcett and Walshaw (2012)](https://doi.org/10.1002/env.2133). There are 3 independent parts to the inference.
1. A Bernoulli (*p**u*) model for whether a given observation exceeds the threshold $u$.
2. A generalised Pareto, GP (*σ**u*,*ξ*), model for the marginal distribution of threshold excesses.
3. The $K$-gaps model for the extremal index $\theta$, based on inter-exceedance times.
For parts 1 and 2 it is necessary to adjust the inferences because we expect that the data will exhibit cluster dependence. This is achieved using the methodology developed in [Chandler and Bate (2007)](https://doi.org/10.1093/biomet/asm015) to produce a log-likelihood that is adjusted for this dependence. This is achieved using the [chandwich package](https://cran.r-project.org/package=chandwich). For part 3, the methodology described in [Süveges and Davison (2010)](https://doi.org/10.1214/09-AOAS292) is used, implemented by the function `kgaps` in the [exdex package](https://cran.r-project.org/package=exdex). The (adjusted) log-likelihoods from parts 1, 2 and 3 are combined to make inferences about return levels.
We illustrate the main functions in `lite` using the `cheeseboro` wind gusts data from the [exdex package](https://cran.r-project.org/package=exdex), which contains hourly wind gust data from each January over the 10-year period 2000-2009.
### Frequentist inference
The function `flite` makes frequentist inferences about $(p_u, \sigma_u, \xi, \theta)$ using maximum likelihood estimation. First, we make inferences about the model parameters.
```{r, echo = FALSE}
got_exdex <- requireNamespace("exdex", quietly = TRUE)
```
```{r freq, eval = got_exdex}
library(lite)
cdata <- exdex::cheeseboro
# Each column of the matrix cdata corresponds to data from a different year
# flite() sets cluster automatically to correspond to column (year)
cfit <- flite(cdata, u = 45, k = 3)
summary(cfit)
```
Then, we make inferences about the 100-year return level, including 95\% confidence intervals. The argument `ny` sets the number of observations per year, which is $31 \times 24 = 744$ for these data.
```{r returnLevels, eval = got_exdex}
rl <- returnLevel(cfit, m = 100, level = 0.95, ny = 31 * 24)
rl
```
### Bayesian inference
The function `blite` performs Bayesian inferences about $(p_u, \sigma_u, \xi, \theta)$, based on a likelihood constructed from the (adjusted) log-likelihoods detailed above. First, we sample from the posterior distribution of the parameters, using the default priors.
```{r seed, echo = FALSE}
set.seed(26012023)
```
```{r Bayes, eval = got_exdex}
cpost <- blite(cdata, u = 45, k = 3, ny = 31 * 24)
summary(cpost)
```
Then, we estimate a 95\% highest predictive density (HPD) interval for the largest value $M_{100}$ to be observed over a future time period of length $100$ years.
```{r predinterval}
predict(cpost, hpd = TRUE, n_years = 100)$short
```
Objects returned from `flite` and `blite` have `plot` methods to summarise graphically, respectively, log-likelihoods and posterior distributions.
### Installation
To get the current released version from CRAN:
```{r installation, eval = FALSE}
install.packages("lite")
```
### Vignettes
See `vignette("lite-1-frequentist", package = "lite")` and `vignette("lite-2-bayesian", package = "lite")`.