Possible improvement to simple re-ordering by interpolating first:

https://onlinelibrary.wiley.com/doi/abs/10.3982/ECTA7880

• i think introducing a class for the covariance matrices would make sampling a lot more user-friendly

• e.g. if we have a temporal covariance matrix then having attributes for the lead time/or spatio-temporal matrices we could have attributes for location and lead time or something along those lines

• at the moment the ordering of the matrix and marginal inputs is v. important in the sampling function, and could be confusing

cov.cor Function name!!!

Think it might be worthwhile only call specific functions we need from each package to avoid conflicts, what do you think?

See end of accepted answer here
https://stackoverflow.com/questions/35750967/r-roxygen2-imported-packages-do-not-appear-in-namespace

I came across this package which seems promising, but a critical part for mass adoption is proper documentation. Pointing people to an Example.R script isn't the right communication strategy. Can you guys please right tutorials on different use cases? Really appreciate it.

• You could link to some resource (papers, blog posts, etc.) that describe more about what/why a user might want or need this package
• The example script doesn't feel like an example, it feels like a work-in-progress. It has a lot of commented code but no explanations of why it is commented, there's no introduction as to what the script is going to do, nor any discussion of almost any kind
• The example script also feels too long, with lots of unnecessary code included
• The files holding your functions are not consistently named. There's MultiQR and MQR, underscores and CamelCase and snakeCase and lowercase.
• There's odd string mangling in read_holiday_ics in DateFeatures.R. I think it will change e.g. "Year Fives" to "Year's".
• It's actually just odd to have date features within this repository. I can see why it would be related to forecasting, but I'd expect that to be a separate package.

We want to be able to:

• store models fit using all data supplied to function (new class?)
• make predictions based on new input data (s3 method?)
• update model based on new input/output data (s3 method?)

Implementation:

• Add option to existing fitting functions to return a list() that contains the [[1]]=MultiQR cv forecasts plus model info (model for each quantile, metadata e.g. quantiles). Model info given new class with name matching fitting function name, e.g. [[2]]="ProbCast_gbm"
• Made sure not to save duplicates of data!!! May require manual removal...
• Predict S3 method for above class: predict(object, newdata,...) return MultiQR
• Update (start with GAM/mgcv and see how it goes...)

Actions

• @ciaran-g Start with gbm: new arguments "keep.model.all=F"
• @jbrowell Strat with GAMs

Too many functions in single script. Need to reorganise into smaller files.

Aim to merge all branches and sort mid-May 2019

Bug in samps_to_scens.R for passing MultiQR objects through the inverse PIT function

Use S3 methods for transforming different model output to multiQR, and for performing the PIT on model classes.

• added fastinvpit function for speed, test with gpd tails

Some distribution families return parameter names outwith the usual "mu", "sigma", "nu", "tau". For example see the BetaLSS() family.

Fixes required for the gamboostLSS_2_multiQR and gamboostLSS_2_PIT functions

To date tails have been limited to [0,1] or fixed user specified max/min.

New functionality:

1. Relative to p50 (allready in multi-qr-mboost branch)
2. Function of p50/covariate
3. User supplied vector of max/min values
4. Extreme value theory (may need additional functions) - Generalised Pareto Distribution?

Change default results to report individual folds, all folds, and test results only, not ALL (folds+test)

Add support for LightGBM as alternative to gbm.
http://jmarkhou.com/lgbqr/

E.g.

gen.trun(par = c(0),family = NO,type="left") produces a new gamlss family "NOtr" and function dNOtr etc, but ppd_gamlss() can not find them...