if you "pull the slack" out of a win probability chart, the length of the resulting line is a measure of the game's excitement. Top 4 games from Week 10:
This flowchart shows my results from simulating NFL playoff overtime, a hot topic since the Super Bowl.
My numbers indicate that there is effectively no advantage between choosing to kick or receive, which aligns with research done by
@bburkeESPN
and
@StatsbyLopez
.
Details⬇️
Much has been made about the divergence between
@FiveThirtyEight
's
#NBA
predictions this year and the corresponding Vegas lines.
For every '21-'22 regular season game, I tested how RAPTOR spreads and win probabilities compared to sportsbooks'.
full post:
if the NFL got rid of overtime and let games end as ties, there’d be way more variety in final record/win% and they wouldn’t need to determine the playoff pool with all those unsatisfying tiebreakers
full post:
I made this map of NYC that scores every block based on how much stuff (bars, restaurants, grocery stores, laundromats, etc.) is on it
full blogpost here:
made with
#rstats
, as always
motivation/implementation/caveats in thread below ⬇️
I'm also attaching this version of the chart with a legend, which is slightly less twitter-friendly. The methodology is:
-Receiving team kicks XP on first possession
-Kicking team goes for two down 7
-Defensive scores not shown (happens about 1% of time)
the irony here is that our distaste for ties doesn’t actually eliminate them — it just shifts them from the game level and into the standings, where they’re much more consequential
three (3) whole friends asked me to pull the slack out of this one so I assume thousands of others feel the same…
second most “exciting” playoff game since 2011! only behind Ravens 38, Broncos 35 in 2012 (Jacoby Jones Hail Mary)
Jan 23, 2022
Final (OT): KC 42, BUF 36
Excitement Index: 7.413
Percentiles (Postseason games):
▪100.0% in 2021 (1 of 10)
▪99.1% since 2011 (2 of 122)
▪100.0% since 2011 for KC (1 of 15)
▪100.0% since 2011 for BUF (1 of 7)
#BillsMafia
#ChiefsKingdom
#BUFvsKC
But I do think that finding such a fundamental (and seemingly easily addressed?) flaw in their predictions makes it harder to buy into their
#NBA
forecasting methodology as a whole.
If your opponent scores 7 on the opening drive, you want to go for 2 because it's a 50% win proposition compared to 40% extending to sudden death where they have the ball first.
While I'm confident in my approach above, I'm sure there are legitimate critiques of the choices I made so I welcome any and all feedback. You can see the code here:
The assumptions shouldn't be too controversial. If you score on first drive, you don't want to go for 2 because then more than half of your opponents' responding TD drives will end up in losses (the 50% of the time you fail your 2-pt, and the ~10% they match and then win later).
Random thought: does anyone ever actually look at the code for
@ben_bot_baldwin
?
How valuable is open source if no one cares to look through the source code?
Or maybe the value is in being able to look if one wanted. idk
That doesn't mean the predictions are useless though. Once you remove the home team bias, a linear blend of sportsbook lines and RAPTOR lines actually slightly outperforms Vegas, at least for moneylines.
@NFLPinnacleBeat
@bburkeESPN
@StatsbyLopez
I agree with that. There are a lot of ways that accounting for team quality will budge these distributions one way or another.
But I find it unlikely that you could shift the probabilities enough to say that there’s a default, obvious choice.
If your defense is gassed then you might consider that first possession more score-heavy than reflected above; if you have one last trick play ready for your 2-pt conversion, maybe you don't mind having to go for 2; and so on
@DSP4150
@benbbaldwin
ha agreed...was hoping I'd get something a little more aligned with the experience. Ties get a big boost from the added overtime plays (and I don't normalize for game length)
The upshot is that both RAPTOR spreads and especially its win probabilities are biased toward the home team, and in particular for home underdogs. Home teams that RAPTOR thinks should win half the time are only winning about 35% of those games
#rstats
@DWAZ73
Sanchez played pretty well in that game. Cold road game against Pitt D and he turned 24-0 halftime blowout into 24-19 game. Needed the ball one more time but D couldn't get stop. Go kick rocks, Jason ✌🏻
The logic between nodes should also be pretty straightforward. Kicking teams down a touchdown will never attempt a field goal, XP are some times unnecessary, sometimes the difference between outcomes, etc.
If OT was eliminated, there’d also be slightly higher correlation between point differential and win% because all those noisy one-score OT games would be “regressed” into half-wins
Simulating was done by stitching together drives from resampled plays in recent seasons using
@nflfastR
data. Constraints were added to make sure teams didn't do anything insane, e.g. kick a field goal down 7, and fourth down decisions were fed into
@benbbaldwin
's model.
The key then is to figure out the transition probabilities between nodes. How do we determine the frequency at which teams will score 0, or 3, or 6-8 points? We can use a combination of empirical results and simulation.
The results illustrate how close to being 50-50 the choice to kick or receive is under reasonable assumptions, which gives coaches the right to have a preference if they feel confident that their particular situation skews the underlying distributions.
The victim in this crash was a friend of mine, Jim Pagels. We were supposed to chat soon to compare our cities' respective bikeshare programs.
Jim was a talented writer whose thoughtful, stats-oriented posts inspired me to try my own hand at blogging.
1/x
Plugging in a blend of these simulations and empirical distributions yields the percentages you see above. Chaining together the probabilities and gathering them at the end-of-game nodes gives nearly a 50% win probability for each team.
For some nodes, simulation isn't necessary -- we have years of "old school" OT games to determine how often each team wins in a sudden death scenario. Other situations have little data (down 3 after the first possession), or none (down a touchdown after the first possession).
new week, same idea -- straightening out the win prob graphs to find the most "exciting" (uncertain, volatile, etc.) games
plz don't yell at me if you didn't actually enjoy these games
Recommender systems ranked:
1. YouTube’s sidebar
2. Spotify’s Discover Weekly
.
.
.
991. My dad sending me videos of tractors he likes
992. Whatever tf Netflix thinks it’s doing
Excited to announce that I'll be presenting my research into NFL playoff overtime on May 16th at the 2024 New York R Conference (
@rstatsai
)
🎟️Tix & Info:
#rstatsnyc
|
#rstats
This flowchart shows my results from simulating NFL playoff overtime, a hot topic since the Super Bowl.
My numbers indicate that there is effectively no advantage between choosing to kick or receive, which aligns with research done by
@bburkeESPN
and
@StatsbyLopez
.
Details⬇️
I refuse to believe I live in a world where every ten minutes someone reports that ChatGPT painted a masterpiece or wrote an original symphony but identifying a half dozen vaguely distorted digits still counts as inimitable human intelligence
@Tucker_TnL
I'm a little skeptical that Vikings still had 30% win prob on Minneapolis Miracle (no timeouts, :10 remaining). ESPN has it much closer to 0%
None of this is meant to denigrate RAPTOR, which was designed to evaluate individual players and not to predict games, or
@FiveThirtyEight
in general -- I've long admired a lot of the work that they put out.
If your opponent scores 7 on the opening drive, you want to go for 2 because it's a 50% win proposition compared to 40% extending to sudden death where they have the ball first.
useful little
#rstats
tip
if you want to create a new ID/index of clusters of data (i.e. consecutive rows with same value in a column), you can simply take a cumulative sum of instances when the lagged value isn't the same:
Last night my friends and I attempted to take part in a timeless American ritual: watching a bunch of NCAA Tournament games at a sports bar.
Unfortunately we stumbled into one of the least competitive 2-hour stretches of basketball in recent March Madness Round 1 history. ⬇️
flight attendant: Are you willing to sit in the emergency exit seat?
me: Sure haha just gotta kick open the hatch and parachute out right?
her: Seriously.
me: 😔 Yes ma'am.
one of the unexpected joys of blogging was paying illustrators for original work.
here are a few of my favorites. DM me if you'd like any of their contact info
me wondering whether I can treat each of the three babies crying behind me on this plane as independent events as I calculate the probability I get any sleep whatsoever
My twitter bio says “always blogging about data,” but in truth I haven’t posted new content in nearly 4 years.
So I’m getting back into it. Semi-regular content going up on . First post, exploring Google's autocorrect functionality, is live!
#rstats
good feature! my friends and I have been competing to find the most obscure answers everyday.
I do wonder if trying to maximize -log(x) would be better than minimizing sum(x) though. my instinct is that a 19% answer and a 1% answer is better than two 10% answers.
🚨 NEW: RARITY SCORE 🚨
After completing your game, you'll be able to see your rarity score on the summary page. The lower, the better! Shout out to
@DSzymborski
@FoolishBB
@JessicaDBrand
for inspiration
Trying to find uphill walking routes shows power of relaxed constraints.
Under strict uphill requirements, I only get a few blocks before hitting local max. But under loosened constraints (elevation can decrease by 1 ft btwn blocks), I make it 50 blocks and finish at global max
when I first got into statistics and forecasting, I read Black Swan and loved it. I also read books like Signal & Noise and Nudge, and loved them too. Turns out the author of the former now spends his days childishly trolling the authors of the latter on Twitter.
what a waste