thegreatwhiteshark / climex Goto Github PK

Since I want to add additional station data to the app the calculation of the return level for all stations will take forever. Also the "minimal number of year" slider is not the most appropriate one for this selection either. So what to do when you just want to compare all stations in a certain area?

I already used such kind of feature in the 'General' tab in order to exclude points of the time series from the fit. It should be possible to use/implement it in the leaflet map as well.

I read https://github.com/theGreatWhiteShark/climex#why-is-this-not-on-cran, but I am very confident that all of the R CMD check issues in this package could be fixed. In particular use of utils::globalVariables() to exclude non-local variables from the global variable check.

R CMD check also reveals a number of other true issues, so even if you don't fix everything or plan to submit this to CRAN I would recommend fixing as many of the check issues as you can.

At least not via devtools::install_github

For now the app is rendered quite nicely on desktops and laptops. But on smartphones it's a different picture.

There's the nice VGAM model using which one can perform non-stationary fitting of the GEV and GP distribution.

I certainly will not implement the VGAM and VGLM methodology myself, but I most probably will add one or two convenience functions linking the VGAM package with climex without the user having to restructure and format her data.

I'm in the process of checking VGAM's performance and consistency right now.

Due to its dependence on the car package via the extRemes package climex can not be installed on R-3.1.3. This is not really nice and I should remove this dependency.

The two files of the shiny app grew organically and became way too big.

It's necessary to separate them into individual files.

The current implementation uses a data.frame consisting of four columns: name, latitude, longitude and altitude. This is not really a nice and intuitive way of structuring the data.

I should have a look at the sp package and figure out a better format.

According to Lee Fawcett & David Walshaw (2007), the calculation of the standard errors for the return level in the GP analysis based on the delta-method (Rao, 1973) is not recommended.

Walshaw (1994) showed the likelihood surface to be "severly" asymmetric. That's why the delta-method relying on the limiting quadratic form of the likelihood surface will produce misleading results.

Instead they suggest to use the profile likelihood described in Venzon & Moolgavkar (1988) or Stuart Coles (2001).

Since I have a decreasing error estimate for higher quantiles when fitting GP functions with negative shape parameter, I clearly have to introduce a better approximation in here.