Monday 10/15/2018 by phishnet


[we'd like to thank Paul Jakus, @paulj, for his second deep dive into the statistice of Phish. - ed.]

Phish fans are famously obsessive about the band, and the setlists on .net reflect that obsession. Setlists on are not simply a list of songs played; instead we are provided with a wealth of information about debuts, bustouts, narrations, and even notation denoting two types of segues. All of this information tries to characterize what happened at a concert (far more than just listing a bunch of songs) with the goal of pointing us to shows of particular interest.

But does this additional information have value? Does knowing the number of “->” segues, or that a song hasn’t been played in over 100 shows, truly correlate with how fans of the band perceive the quality of a given show? Can we look at an intriguing setlist from 10 or 15 years ago (or look at today’s setlists a decade from now) and use its content to determine whether or not to listen to the show? That is, do the elements of a Phish setlist relate to how we, the members of, rate Phish shows? show ratings are somewhat problematic from a statistical standpoint, but let’s put that aside and simply say that show ratings, for good or bad, reflect only the opinions of .netters. (See the third paragraph of this blogpost for problems regarding .net show ratings.)

I gathered a bunch of setlist data for 359 3.0 era shows, from the Hampton Reunion run through the 2017/2018 NYE run, and estimated a regression model linking setlist notes to show ratings. PM me if you want the detailed model, but let’s just highlight the things that were statistically related to the show rating on

Setlist Elements and Show Ratings

Setlist Notes Related to Show Rating

Setlist Notes Not Related to Show Rating

Number of Songs per Set (−)

Debut songs

Number of “>” segues (+)

Vacuum solos

Number of “->” segues (+)

Guest artists

Average song gap (+)

Narration songs (+)

Festival Shows (+)

Three set shows (+)

Shows on a Weekend (+)

Time Trend (+)

The “±” sign in parentheses is the direction of the effect on the show’s rating. For example, the more songs that are played the less we like it (negative) and the more “->” segues that are played the more we like it (positive).

Here’s a graph that uses the statistical model to depict the actual show rating on the vertical axis and the predicted show rating (“fitted value”) on the horizontal axis. A perfect model would have its predicted value exactly equal to actual value, so that every dot would lie along the dashed line. Of course, the model was not perfect (it explained just under 40% of the variation in show rating) but you can see that, broadly speaking, the model predicts a high show rating when a show actually was rated highly, and vice versa.

Now you might think that explaining only about 40% of variation in show rating is not very good but I was pleasantly surprised: I didn’t think it would explain much more than 25% or so. The setlist information does a decent job of identifying the characteristics of a good Phish show.

But can a statistical model based on 2009-2017 data predict how .netters would rate shows in 2018? Let’s take a look at our actual ratings for the Summer 2018 tour, and compare them to what the statistical model would predict, given the 2018 setlists:

Actual .net Ratings versus Model Predictions


Actual .net Rating

Model Prediction

Tahoe N1



Tahoe N2



Gorge N1



Gorge N2



Gorge N3









Forum N1



Forum N2



Del Valle



Alpharetta N1



Alpharetta N2



Alpharetta N3



Camden N1



Camden N2












Dick’s N1



Dick’s N2



Dick’s N3



As it turns out the model didn’t perform all that badly. Out of 21 shows from Summer 2018, the model predicted 15 to within 10% of its actual value. Sometimes a graph is better than raw data, so here’s the absolute percentage of prediction error plotted against the actual show rating:

As you can see, the model predicts pretty darn well for shows that have a show rating of about 3.8 or higher, but that the errors get large with the lowest rated shows. One show, MPP N1 (at the top left of the graph), is off by a whopping 25%.

In fact, MPP N1 caught my attention the day after it was played: I had missed the webcast and logged onto .net the next morning to check out the setlist, which looked fantastic. An eight-song first set with "Blaze On," "Stash," and "46 Days"? A five-song second set filled with heavy-hitters?! A four song encore!?! This must have been an incredible show, right?

Umm, no, apparently not (we rate that show as a 3.49).

So, though the setlists satisfy their primary job of pointing us to good shows, MPP N1 helps highlight the shortcomings of any statistical model. The model can only measure what’s on paper, and can’t really measure the magic of an improvisational performance. We can analyze setlists all we want, but we still gotta to listen to the show.

If you liked this blog post, one way you could "like" it is to make a donation to The Mockingbird Foundation, the sponsor of Support music education for children, and you just might change the world.


, comment by 3PointOhKeePa
3PointOhKeePa Great work! I love any Phish related statistical analysis, and this really does make you step back and think of how we view shows. It's funny, because I personally love MPP N1, and would probably agree with the model rating, but then that goes back to how subjective shows are and why we really should take the ratings with a grain of salt when looking for shows to listen to.
, comment by MultiBeastMode
MultiBeastMode This is great to see! Funny thing is that for the past 6 months or so, I'd been thinking that an analysis of the ">" and "->" frequency, plus the number of songs in a show and the length of a show (would be harder to get that data), would be fun to do. I know some great set 2s are on the shorter side time-wise. But that's purely anecdotal and might not be supported by data. Good job on doing something that I'd probably never have gotten off of my ass to do.
, comment by paulj
paulj @MultiBeastMode said:
I'd been thinking that an analysis of the ">" and "->" frequency, plus the number of songs in a show and the length of a show (would be harder to get that data), would be fun to do.
Yeah, for the past several years I've been of the opinion that the best "two-set plus encore" Phish shows were 2.5 hours long and had 17 or 18 songs. I tried to put a non-linear (quadratic) term for # of songs, but the linear term was best. Instead, it's pace and flow (the ">" and "->" segues) matter a lot, plus things that I can't measure in the model: song choice and set placement, and improvisational magic.

I'm actually rather pleased that the model works to some degree (setlists do convey good information) but that they don't capture everything.
, comment by NoHayBanda
NoHayBanda Nice number crunching. It makes sense if a show looks unique more people are likely to listen to it, like it, and vote on it. Easy to look at MPP2 and say "I gotta hear this!" where something like The Forum N2 doesn't invoke the same reaction. The number of voters in 2018 range from 200 - 500 and its easy to see the correlation between high rating and most votes.
, comment by positivevibes
positivevibes Fantastic work! There is much that data science can give us into the world of Phish. The one thing that can make a show lower rated than the algo would provide is something that is hard to glean from a setlist- flubs. I think in some of the shows that are rated higher than the phans rating is the flubbing and possibly meandering jamming and indecisiveness. I rate shows based on 1. quality of playing 2. variety of songs 3. flow (very important) 4. synchronicity (jamming and songs-wise). Most of this is intangible and thus I think is what makes the discrepancy in the algo and phans opinions.
, comment by zothound
zothound @3PointOhKeePa said:
Great work! I love any Phish related statistical analysis, and this really does make you step back and think of how we view shows. It's funny, because I personally love MPP N1, and would probably agree with the model rating, but then that goes back to how subjective shows are and why we really should take the ratings with a grain of salt when lookin

for shows to listen to.
Curious of your thoughts on N1 mpp. Granted, I've only seen phish 14 times but I was legitimately bored in set 1. It was the one first time I was even close to bored at a phish show. See? I'm a noob.

Set 2 was good, the Sand was average but nice. Mercury ghost the highlight. Also loved the Martian monster version.

Set 1 though... Yikes.
, comment by lysergic
lysergic The fact that the model can explain 40% of variation in show rating is incredible. Think about it. That means 40% of the perceived quality of the show (as evaluated by .netters) can be explained purely by looking at the show on paper. You don't have to hear a single second. We are talking about a band famous for improvisation and fans who can tirelessly analyze the details distinguishing different live versions of the same song. More than that, we are talking about music and all of the emotions it invokes. To state the obvious: looking at a setlist on paper and listening to a show are completely different phenomena. Presumably the show ratings are based on the latter, not the former.

Looking at the model output, you can tell what the ideal 3.0 show looks like on paper:
--three sets, with only a few songs per set
--lots of segues
--major bustouts
--narration songs
--played on a weekend at a festival

...sounds like a good time to me!

Also the time trend indicates that either recent shows have biased ratings, or the band is improving over the course of 3.0. Probably both are true.
, comment by Hendrix_Phishinfloyd
Hendrix_Phishinfloyd :: Instantly changes all show ratings to 5-stars ::
, comment by paulj
paulj @NoHayBanda said:
... its easy to see the correlation between high rating and most votes.
Yes, but how should one deal with this? We're using average rating, so as long as the central tendency of the distribution is constant, it shouldn't matter how many ratings it gets. On the other hand, the # of ratings could be a proxy for the "intangibles" of a show. More people rating ==> Better show. I'm sure the effect is positive and would improve the predictive capability of the model, but it's hard for me to justify this as something that happened during the show (which is what all the other variables measure.)

Oh, the # of raters would be lower the more recent the show...and show ratings also fall over time. So those add more complications.
, comment by paulj
paulj @Hendrix_Phishinfloyd said:
:: Instantly changes all show ratings to 5-stars ::
Ha! After gathering the data before Summer 18 and seeing the average show rating (3.96), I went back and re-rated a bunch of shows after realizing I had been too promiscuous in my use of 4's and 5's. I tend to rate only those shows I've been to, and now I have only three shows rated 5 stars: 11/22/97, 12/31/99, and 10/31/14.
, comment by Pbjams
Pbjams I left MPP N1 feeling like it was an exceptional show, better than seeing Raleigh the night prior (which was fun but not as crisp with a poor sound setup near the front), and still scratch my head at the rating. The band had a few near perfect jams and suspect the reviews were based on more on playlist (which doesn't look that great) and less on performance.

If you love rarities or throwbacks, MPP n1 is not the show for you but KyleIndeed's review of it captured how solid a large part of that show actually was, worthy of the model projection. Neat study.
, comment by jonesgator
jonesgator I love this! I think a model predicting a setlist should be next (or maybe we already have one)! One thing about this model - there is going to be some endogeneity due to omitted variable bias, caused by our inability to measure quality of playing. I bet if you had some kind of metric for quality your R2 would be above 90.
You must be logged in to post a comment.
Support & Mbird
Fun with Setlists
Check our Phish setlists and sideshow setlists!
Phish News
Subscribe to Phish-News for exclusive info while on tour! is a non-commercial project run by Phish fans and for Phish fans under the auspices of the all-volunteer, non-profit Mockingbird Foundation.

This project serves to compile, preserve, and protect encyclopedic information about Phish and their music.

Credits | Terms Of Use | Legal | DMCA

© 1990-2023  The Mockingbird Foundation, Inc. | Hosted by Linode