|
Forever_Peace posted:That is exactly one of the biggest questions I had, too. Like, 5.6 yards per carry is a lot, but with 93 carries, could that just have been due to random chance? What if it was 120 carries? 150? When would we start to believe he's a stud? If you're going to delve into dealing with the statistics of small sample sizes, you would be well-served to read this article: http://press.princeton.edu/chapters/s8863.pdf In particular, figure 1.3 is relevant to the issue of 'judging performance of many RBs of varying sample size', although not a perfect comparison (of course!).
|
#
¿
Feb 10, 2016 00:35
|
|
|
|
| # ¿ Apr 29, 2024 04:34 |
|
|
Forever_Peace posted:
I haven't read your code, but I'm wondering about how you can rigorously evaluate wildly different quality of offensive line play? There's a proposition (quoted above) that it will shift the cumulative distribution but not affect its shape, but has this been tested/validated? Is it even possible to validate with the data on hand? That seems the most difficult part of comparing players cross teams. When a lineman pancakes a defender, the run breaks for an random number of additional free yards. That would change the shape of the cumulative distribution, not just shift it. It seems really problematic for this kind of player analysis. Anyway, the thread is really fun and I do appreciate the effort put forth into doing the best you can do with the data on-hand.
|
|
#
¿
Feb 13, 2016 18:11
|
|
“Yeah, but take a look at where you actually start a run, a few yards behind the line of scrimmage. The opposite is true. Every step you take is, on average, a step closer to being tackled.”
|
Forever_Peace posted:The inability to capture and express these influences would be very hard for predictions, yes, but I've tried to be very clear that I'm mostly constraining myself to retrospective explorations so far. This is important because in a retrospective analysis, there is no "novel". We don't need to try to project how a player should perform given some new offensive line - we can simply interpret a player's distribution as already reflecting some combination of that player and his surrounding cast (including offensive line, quarterback, and playcallers). This is a well phrased and very important distinction. I imagine many people are interesting in using player comps to predict future performance, whether for fantasy football or discussing draft/FA signings of their favorite team. Prediction is really difficult. Thanks for the reply, fun thread. Maybe I'll peek at your code and see if it's easily pythonized 
|
|
#
¿
Feb 15, 2016 17:07
|
|
|
John Brown posted:E: Actually, I have a request. Would you mind doing an analysis on Melvin Gordon in the same way you hit on TJ Yeldon? It's thrown around a lot in these threads that Gordon sucks and I'd like to know whether or not it's too early to make that assessment. According to NFL.com he had 184car, 641yds, a 3.5 YPC average, and no TDs. Just read his latest big post in detail. Specifically, quote:When you see somebody throwing around yards per carry with small sample sizes, say: 150 < 184 < 200 and 3.5 < 3.81, therefore the entirety of the San Diego Chargers rushing game while Melvin Gordon was the chosen ball carrier was distinctly below NFL average. However, we don't know if Melvin will be below average in the future or not.
|
|
#
¿
Mar 1, 2016 00:37
|
|
|
The Chargers' line was bad going into the season and became worse with injuries. That explains most of last year's performance for Woodhead and Gordon. Keep an eye on draft/FA for SD OLine.
|
|
#
¿
Mar 13, 2016 02:09
|
|
|
Next step to developing a predictive model should be trying to separate RB performance from the surrounding cast, but I don't know if you have the database for that on hand. Going to give it a shot? Or perhaps a confidence based model that can quantify similarity year-to-year in surrounding cast, and then forecast RB performance with some variable margins for error, or distributions of results? Basically, you have dozens of readers wanting to know who to draft in fantasy this season (that's okay to talk about in broad generalities, here, I hope?). But you already knew that 
|
|
#
¿
May 4, 2016 03:17
|
|
|
Forever_Peace posted:Bottom Line I think your disclaimer at the end deserves more emphasis. From your own CH4 ( http://forums.somethingawful.com/showthread.php?threadid=3763012&userid=0&perpage=40&pagenumber=3#post456827392 ), the performance of either player "could easily be explained by the random variance of a perfectly average running back." We simply don't know which is the better runner. One thing we know for sure, though, is that Freeman is good at catching the ball.
|
|
#
¿
May 30, 2016 18:55
|
|
|
Going into week 17, I'm going to revisit some things I commented on much earlier in this thread: On Melvin Gordon -- pmchem posted:Just read his latest big post in detail. Specifically, Forever_Peace posted:Quick Hits: What the gently caress is up with Melvin Gordon? That above quote is heavily edited for space, no disrepect, but I wanted to include points that reflected both sides to the analysis. Yeah, the SD running game was real bad in 2015. That appeared to affect Gordon and Woodhead. But, even knowing that, Gordon gets hit bad in the final analysis -- which is phrased as predictive analysis. The JAG comment was probably most fitting to describe 2016. His YPC in 2016 was, indeed, explainable by that of a league average running back (still much better than 2015!). It's just so dang hard to separate out the effects of play-calling, health, and line play from individual performance sometimes. Re: Freeman vs. Coleman -- pmchem posted:I think your disclaimer at the end deserves more emphasis. From your own CH4 ( http://forums.somethingawful.com/showthread.php?threadid=3763012&userid=0&perpage=40&pagenumber=3#post456827392 ), the performance of either player "could easily be explained by the random variance of a perfectly average running back." We simply don't know which is the better runner. One thing we know for sure, though, is that Freeman is good at catching the ball. Forever_Peace posted:Yep, that's totally fair! 87 carries is really not a lot. Based on record alone, it's probably too soon to say. Freeman turned out noticeably better than Coleman this year in YPC, but we face the same problem as last year: Coleman still suffers from sample size issues. ATL offense w/ Coleman could be explained by a league average back this year, but not w/ Freeman. I haven't examined the distribution for either. Amusingly, if you compare their aggregated 2015-2016 stats, their YPC is 4.25 vs 4.41. Basically indistinguishable, especially given sample size disparity. Anyway, I hope this thread leads to more fun, novel analysis in the upcoming offseason. Maybe with new data sources! Who knows? Graphs are great.
|
|
#
¿
Dec 30, 2016 04:57
|
|
|
Forever_Peace posted:
Good stuff. I see you added the log equation for surprisal as compared to the draft chapter. Could you also add the explicit formula for game probability? It would save some math nerds a trip to the R code to verify that the math you (presumably) implemented in the code matches the reader's interpretation of the verbal description given in the "whole-game probability" section. I still want to see how this compares to some other historical data on advanced metrics or fantasy ranks, would be really interesting. e.g. surprisal ratings using data for 2010-2015, ranking those players, compare to fantasy rankings pre-season 2016, compare to other advanced metrics' ranks of players using ~2015 era data. Could do this for a whole series of years, of course. The questions being: is this merely an interesting statistic that yes, confirms our ideas about particular examples such as JCharles? Does it have predictive power greater than some other advanced metric? Does it confirm intuition from other sources? An interesting tangent to that would be: how far does a player need to differentiate himself in one (or multiple) seasons on the surprisal metric in order to get into the HOF? It could be a necessary, if not sufficient, condition for HOF.
|
|
#
¿
Apr 27, 2017 22:49
|
|
|
Leperflesh posted:Fantasy rankings are heavily skewed (or should be) by both receptions and scoring, whereas this surprise rating is entirely focused on yardage (which also means it's still underrating goal-line backs and under-valuing x-and-goal carries). I agree, but it would be interesting. And let's face it: lots of these types of analyses are motivated by the fantasy football community. I mean, unless FP is gunning to get hired as a quant in some front office, his audience has immense overlap with fantasy football players.
|
|
#
¿
May 1, 2017 00:00
|
|
|
I have a NFL data science question and this seems like the most appropriate place. Let's talk raw data sources. Ground Control uses nfldb. As far as I'm aware (and I might be wrong), if a player does not collect stats in a particular game, nfldb does not differentiate between: (1) player was active, but did not play, (2) player was inactive due to injury, (3) player was suspended. I'm also not sure what it does with players who not on a NFL roster for a given week but collect stats other weeks; presumably it has week-by-week NFL roster status for each player. Is there an easy source for data types 1-3? Especially #2+#3? Preferably in a way that can be easily imported via a python interface such as nfldb? I am looking to do a little machine learning, but missing that data would make the effort pointless. Manually entering the data would be prohibitive on my human-time.
|
|
#
¿
Jul 29, 2017 23:32
|
|
|
Ghost of Reagan Past posted:This is actually the hardest part of doing data science. yeah, which is why I'm asking around. I looked at PFR and their weekly data seems limited to starters and participants. No information about deactivation due to injury vs suspension.
|
|
#
¿
Aug 6, 2017 15:56
|
|
|
Forever_Peace posted:I don't know either but if you find a source let me know. Well if you find a source of that data, I'll throw some of my ML skills at some things. Need the data, though.
|
|
#
¿
Aug 15, 2017 23:04
|
|
|
Forever_Peace posted:Some other players 2+ SD above average: The fact that "the Law Firm" BJGE makes that list should only reinforce the fact that it's really the rushing offense that is being judged here, not the individual RB. While there may be a high correlation between the two, false positives (and negatives) will sneak in. BJGE is a career 3.9 YPC guy who managed a high TD/carry rate due to playing on the Patriots a few years.
|
|
#
¿
Aug 27, 2017 23:26
|
|
|
Well, without being able to peer into your formula and see how much each particular factor counts towards your EPA, it's hard to frame an argument. I can only discuss by example. So, example: David Johnson has 9 career fumbles compared to BJGE's 5, and DJ has those fumbles on a little more than half the career touches of BJGE. I'd take DJ over BJGE on my football team any day of the week and twice on Sunday. They were rated similarly in your SD metric. BJGE, on the other hand, took years to earn a one-dimensional / GLB rushing job with the Pats. Their (arguably best front office in the NFL) let him walk in FA. He had one good season on a 3-year contract with Cincy, and washed out of the league before age 30, despite not having any major injuries. He's Just A Guy and sometimes fumble rates are just luck. I will never be convinced BJGE is 2 standard deviations above most other RB's. He was in the right spot, at the right time, with the right stocky, power rushing frame to fall forward for the Pats.
|
|
#
¿
Aug 28, 2017 00:05
|
|
|
Forever_Peace posted:EPA is much more sensitive to playcalling and situational factors than other metrics I've worked with. It really reflects players that have been put in a situation to be successful. It's real late after a 20-team draft and game of thrones, so I'm quoting just this part. Yes, I agree with the quoted part! I like your analysis and efforts. But they are often framed around individual players, and the casual reader may read it as an absolute metric of individual player value. There is so much going into it reflecting the rest of the team around the player, the playcalling, and the general offensive efficiency of the team that is extremely difficult to disentangle. You put BJGE on the 2010 Pats, he has a great EPA. You put him on I dunno, the 2010 Cardinals, he's Beanie Wells, nobody remembers him and his EPA is uninteresting. The trick is identifying (algorithmically) the talents that transcend their situation. Charles might be one. Barrie Sanders was one.
|
|
#
¿
Aug 28, 2017 05:47
|
|
|
|
| # ¿ Apr 29, 2024 04:34 |
|
|
I just saw that the nfldb and nflgame github repos were marked unmaintained by the owner. So sad. Did anyone take over on a new fork, or is there a good alternative free, python-friendly nfl data source now? There’s a couple kaggle datasets of questionable quality.
|
|
#
¿
Mar 10, 2018 04:01
|
|