TODO.Rmd

---
output: html_document
---

# TODO(?)

+ [ ] Prefix main functions with `rapm_` (instead of `auto_`).
(This could be useful if additional functionality is added to the project,
such as for calculating team strength.) If this is done, not sure if same
should be done with "one-time" functions (e.g. `combine*()`, `download*()`, etc.).
+ [x] Check that filtering for `poss_min` (and other criteria) is done correctly.
    + I've implemented two versions. One version removes entire lineups
    where any 1 player does not meet the criteria (probably more correct, if
    any kind of filtering is going to be done, although the corresponding
    offensive/defensive lineup should probably also be removed),
    while the other removes individual players who meet ther criteria
    (which seems "less correct", but produces better results(?), anecdotally).
+ [x] Try `intercept = FALSE` with models. 
    + It returns weird results.
+ [x] Try `standardize = FALSE` with models.
    + Seems to generate more accurate results in terms of ranks, but the coefficient
    values are off (in terms of magnitude).
+ [x] Try `scale = TRUE`, where the dummy variables (in the stint matrix) are
normalized according to number of possessions (relative to the maximum number of
possessions played by any stint).
    + This produced some very strange/wrong results.
+ [x] Try negative dummy variables for defense vs. negative response variable for defense. 
    + No difference.
+ [x] Try `collapse = TRUE/FALSE` (i.e. summarising by stints vs. not).
    + `collapse = TRUE` seems to generate more reasonable results (although 
    should have tested `ppp` as response instead of `pp100p` when `collapse = FALSE`)

+ [x] Check/emulate how `{glmnet}` is used at:
    + https://github.com/dgrtwo/data-screencasts/blob/3e1e3d0e14abdff00ae1493d4273e8f3b49a0a77/medium-datasci.Rmd
    + https://github.com/juliasilge/blog_by_hugo/blob/f847b7836bfcedc396e7e82867c4601ae8357853/content/blog/2018/2018-12-24-tidy-text-classification.Rmd
+ [ ] Re-do calculations to follow code shown at
    + https://squared2020.com/2017/09/18/deep-dive-on-regularized-adjusted-plus-minus-i-introductory-example/
    + https://squared2020.com/2017/09/18/deep-dive-on-regularized-adjusted-plus-minus-ii-basic-application-to-2017-nba-data-with-r/.
    + https://squared2020.com/2018/12/24/regularized-adjusted-plus-minus-part-iii-what-had-really-happened-was/
        + [x] Note: I'm 99% confident that I'm not making the additive rating mistake identified in the article
        (because I calculate `pp100poss` AFTER summing `pts` across all possessions `n`).
        + [ ] Maybe try combining offensive and defensive matrices(?), although Jacobs seems to take opinions on both sides
        of the argument of whether or not this should be done.
+ [ ] Re-do `clean` and `reshape` functions with `{data.table}` (and use S3 methods to differentiate with `{tibble}` methods).
+ [ ] Compare calculations with those at http://nbashotcharts.com/ (e.g. http://nbashotcharts.com/rapm?id=1113190703 for 2017 season)
+ [ ] Create function to automatically plot ORAPM vs. DRAPM. (Code/logic is already in
one of the `.Rmd` files.)
+ [ ] LOW PRIORITY: Create function to regenerate profile html widgets.
(Code/logic already exists somewhere in the project.)
+ [x] Move filtering function to last step before fitting (so that `reshape` step
can possibly be skipped).
+ [ ] Evaluate variance of coefficients. (This doesn't seem to come directly
with `broom::tidy.glmnet()`, so it doesn't seem trivial.)
+ [ ] Add box-score plus-minus (i.e. BPM) model and use it to stabilize RAPM models.
    + Possibly can just use `players_summary_nbastatr` data and run a regression.
    Need to consult "the literature" to figure out which features to use as predictors,
    as well as to figure out how to "add" the results of the model to the traditional
    RAPM model.