Skip to content
manuhuthPublic

About

Likelihood-based methods for model fitting, assessment and prediction analysis of some convolution-closed count time series model are provided. The marginal distribution can be Poisson or Generalized Poisson. Regression effects can be modelled via time varying innovation rates.

Resources

Stars

4 stars

Watchers

1 watching

Forks

Repository files navigation

coconots

R-CMD-check julia-tests pkgdown CRAN status Codecov test coverage

📖 Documentation: https://manuhuth.github.io/coconots/

Functions to analyse time series consisting of low counts are provided. The focus in the current version is on practical models that can capture first and higher-order dependence based on the work of Joe (1996). Both equidispersed and overdispersed marginal distributions of data can be modelled. Regression effects can be included. Fast and efficient procedures for likelihood based inference and probabilistic forecasting are provided as well as useful tools for model validation and diagnostics.

Overview of the functionality of coconots

Details

cocoSim() simulates convolution-closed count time series models and cocoReg() fits them. The S3 method predict() computes k-step ahead forecasting distributions for a fitted model. cocoBoot(), cocoPit(), cocoScore(), cocoSoc() and cocoResid() provide routines for model assessment. The main usage of the package is illustrated in the help page of cocoReg(). Further examples are given in the help pages of the other functions.

By default, the functions use a C++ implementation via Rcpp. Users with a Julia installation can instead run them in Julia by setting the argument julia = TRUE. This is most useful for cocoReg(), where the likelihood is evaluated numerically: the Julia backend uses automatic differentiation to obtain exact gradients, which can improve numerical stability and speed of convergence for complex models.

Both backends gave qualitatively similar results in our tests. Rcpp is therefore the default, so that the package works without Julia.

Model

Definition of the model

Installation

The stable version of coconots is available from CRAN:

install.packages("coconots")

The development version can be installed from GitHub:

devtools::install_github("manuhuth/coconots")

Example using the Rcpp implementation

The Rcpp implementation is the default. The regression results can be summarized using the summary() function.

library(coconots)
length <- 500

pars <- c(1, 0.4, 0.2)
set.seed(12345)
data <- cocoSim(order = 1, type = "GP", par = pars,
                length = length)
coco <- cocoReg(order = 1, type = "GP", data = data)
summary(coco)
#> Coefficients:
#>          Estimate   Std..Error         t
#> lambda     1.0546       0.0930   11.3368
#> alpha      0.4202       0.0373   11.2541
#> eta        0.1609       0.0329    4.8961
#> 
#> Type: GP
#> Order: 1
#> 
#> Log-likelihood: -885.012
#> AIC: 1776.024
#> BIC: 1788.6679

Example using the Julia implementation

The Julia functions require a working Julia installation. installJuliaPackages() installs the Julia package Coconots.jl (version 0.2.0 or higher) and updates an older installed version. It needs to be run once.

coconots::installJuliaPackages()
#> Starting Julia ...

The Julia implementation is used by setting the argument julia = TRUE. The seed of the Julia random number generator is set via julia_seed, since set.seed() does not affect Julia.

library(coconots)
length <- 500

pars <- c(1, 0.3, 0.01, 0.2)
data <- cocoSim(order = 2, type = "Poisson", par = pars, length = length,
                julia = TRUE, julia_seed = 123)
coco <- cocoReg(order = 2, type = "Poisson", data = data, julia = TRUE)

Model assessment

The package provides several tools for model assessment. A model fitted with the Rcpp implementation can also be assessed with the Julia implementation. For this, the fit must be created with julia_installed = TRUE, which stores a Julia-compatible copy of the fit. This is not necessary if the model was fitted with julia = TRUE.

library(coconots)
length <- 500

pars <- c(1, 0.4)
set.seed(12345)
data <- cocoSim(order = 1, type = "Poisson", par = pars, length = length)
coco <- cocoReg(order = 1, type = "Poisson", data = data,
                julia_installed=TRUE)
(scores <- cocoScore(coco, julia = TRUE))
#> $log.score
#> [1] 1.501935
#> 
#> $quad.score
#> [1] -0.2532048
#> 
#> $rps.score
#> [1] 0.6197608
#> 
#> $aic
#> [1] 1502.931
#> 
#> $bic
#> [1] 1511.36
pit <- cocoPit(coco, julia = TRUE)
plot(pit)
#> Warning: `aes_string()` was deprecated in ggplot2 3.0.0.
#> ℹ Please use tidy evaluation idioms with `aes()`.
#> ℹ See also `vignette("ggplot2-in-packages")` for more information.
#> ℹ The deprecated feature was likely used in the coconots package.
#>   Please report the issue at <https://github.com/manuhuth/coconots/issues>.
#> This warning is displayed once per session.
#> Call `lifecycle::last_lifecycle_warnings()` to see where this warning was
#> generated.

boot <- cocoBoot(coco, julia = TRUE)
plot(boot)

forecast <- predict(coco, k=3, julia = TRUE)
plot(forecast) #plot all k forecasts

plot(forecast[[1]]) #plot density of one-step ahead forecast

resid <- cocoResid(coco)
plot(resid)

Compare models

cocoSoc() fits the Poisson and generalized Poisson models of order 1 and 2 to the same data and tabulates the scoring rules for model selection.

library(coconots)
length <- 500

pars <- c(1.3, 0.25, 0.03, 0.2, 0.3)
set.seed(12345)
data <- cocoSim(order = 2, type = "GP", par = pars, length = length)
soc <- cocoSoc(data, julia = TRUE)
#> [1] "Computing PAR1 model."
#> [1] "Computing PAR2 model."
#> [1] "Computing GP1 model."
#> [1] "Computing GP2 model."
summary(soc)
#>            Logarithmic Score   Quadratic Score   Ranked Probability Score
#> 1   PAR1            2.265970        -0.1221436                   1.302572
#> 2   PAR2            2.264753        -0.1223451                   1.301770
#> 3    GP1            2.164318        -0.1363304                   1.263041
#> 4    GP2            2.161843        -0.1364925                   1.261362

About

Likelihood-based methods for model fitting, assessment and prediction analysis of some convolution-closed count time series model are provided. The marginal distribution can be Poisson or Generalized Poisson. Regression effects can be modelled via time varying innovation rates.

Resources

Stars

4 stars

Watchers

1 watching

Forks

Releases

Packages

Used by

Contributors

Languages