I feel like we may have enough data on this site to meaningfully break programming languages into classes.

What I mean by classes is akin to Go's ranking system:

If a player can win 90% of even games against a 2 kyu player, the AGA believes he or she is 1.33 ranks higher, the EGF believes (s)he is 2.42 ranks higher, and the IGS believes (s)he is more than 3 ranks higher.

We may need to break the competition down into tags (e.g., a puzzle may have a different ranking system than a puzzle (cough, Bubblegum).

The use case for this would be to more reasonably compare your success - If a Python answer beats one in J, the author should be pretty proud. If Pyth beats Java by just a bit, the Java author should also be pretty happy with themselves (or the Pyth author ashamed).

Would anyone be interested in this endeavor?

  • 1
    \$\begingroup\$ It's an interesting idea, but I think most of us have already formed subjective opinions; for example, I consider Ruby, Python, Perl, etc. in a higher and fairly well-defined "class" than Java, Rust, etc. Nevertheless, if you do attempt something like this, I'd be very curious to see the results! (I have a language-detection script floating around somewhere, in a Github Gist I think, that could come in handy.) \$\endgroup\$ – Doorknob Mar 25 '16 at 16:58
  • 7
    \$\begingroup\$ I think there are two (broad) schools of thought on what "success" means here in golf terms. One is getting the overall shortest by choosing the appropriate language for each task (choosing your tools). The other is finding the shortest given a certain tool (language). While I like to beat up on C# with Java as much as the next guy, it's not a true comparison across all golfs, because the real difference is (mainly) which builtins are there and how they're named. So I'm not sure I see how ranking each would help people in either school of thought judge their success. \$\endgroup\$ – Geobits Mar 25 '16 at 17:03
  • \$\begingroup\$ To echo what Geobits said -- defining the tasks that make up the metric is going to be extremely difficult and subject to much debate. Many languages are very close on the Hello World challenge, simply because "Hello, World!" (with or without the trailing ") is a valid program in dozens of languages. Or ones where a particular language has a clear built-in that leapfrogs it ahead of where it would "normally" sit. Challenges like that would skew the numbers incorrectly. Not that I'm against the idea, it just needs a lot of rigor. \$\endgroup\$ – AdmBorkBork Mar 25 '16 at 17:51
  • \$\begingroup\$ @TimmyD Yes - and I have no statistical background, so I'd like to get someone with more stats knowledge on board before pursuing it. \$\endgroup\$ – Not that Charles Mar 25 '16 at 17:53
  • 7
    \$\begingroup\$ I honestly have no clue why this has been consistently downvoted so far. While the commenters have a point that it's going to be hard to figure out how to get the most meaningful data, I'm pretty sure that across all our code golfs we can find significant trends easily enough. It may be hard to determine with any certainty whether C# or Java is better for golfing, but I'm sure we can detect some fuzzy ranking. +1, I like the idea of trying to mine some data from our content. \$\endgroup\$ – Martin Ender Mar 25 '16 at 18:59
  • 3
    \$\begingroup\$ @MartinBüttner, it's probably suspicion about motives. I'm fighting the temptation to downvote by reminding myself that we should assume good faith, but the context of previous meta questions by people who were pushing handicap systems does inevitably affect the subtext I read into this one. \$\endgroup\$ – Peter Taylor Mar 25 '16 at 19:20
  • 2
    \$\begingroup\$ @PeterTaylor That's a good point, but I'd rather downvote the post making use of such a handicap system than an attempt to gather some interesting data from our challenges. \$\endgroup\$ – Martin Ender Mar 25 '16 at 19:28
  • 1
    \$\begingroup\$ @PeterTaylor I am not very active in meta and haven't read that thread. I don't have motives aside from what is presented. \$\endgroup\$ – Not that Charles Mar 25 '16 at 20:01
  • 12
    \$\begingroup\$ @PeterTaylor Also, for the record, I think a handicap system is a terrible idea. \$\endgroup\$ – Not that Charles Mar 25 '16 at 20:01
  • \$\begingroup\$ Another reason is that it would be nice to see a language competition - some sort of aggregate of all the work folks have done to abuse languages. Where does Perl beat Ruby? Does D beat C? Where does TI-BASIC beat Mathematica? \$\endgroup\$ – Not that Charles Mar 25 '16 at 20:38
  • 1
    \$\begingroup\$ I downvoted because, not only would this be very difficult and many people would disagree with the classes due to the subjectivity involved, but defining language classes opens the door to attempts to impose handicaps, which I am strongly against. \$\endgroup\$ – Mego Mar 25 '16 at 21:59
  • \$\begingroup\$ We already have some data in response to this code challenge. \$\endgroup\$ – xnor Mar 25 '16 at 23:13
  • 9
    \$\begingroup\$ I don't see why people are so opposed to this. Everyone knows that some languages are golfier are other. Most golfers already have a vague sense of the ordering, and surely actual statistics are better. \$\endgroup\$ – xnor Mar 26 '16 at 4:29
  • 5
    \$\begingroup\$ Seems like a fine idea to me. While it shouldn't be considered definitive, it'll probably be interesting. \$\endgroup\$ – isaacg Mar 26 '16 at 7:27
  • \$\begingroup\$ @xnor Like I mentioned before, I don't think it's a worthwhile endeavor. The lines are too fuzzy to be able to create clear-cut classes, and there are lots of issues that skew the data such that the classes would not be meaningful (as I commented on Alex's answer). \$\endgroup\$ – Mego Mar 29 '16 at 1:18

Elo ratings, treating pairs of submissions as games.

I'm using the Elo system here.

Essentially, the solution is find the best solutions in each language to each challenge, and then treat each pair of such solutions as a match between those two languages, with the winner being the shorter submission.

Languages with less than 10 entries are discarded, and languages have a starting rating of 1000.

One drawback of this system is that ELO favors more recent entries. To compensate, I used a relatively low K value, which makes updates slower, and so old answer count for more.

The data is from this query, courtesy of Alex A. To run the code, put the two code blocks below in files with the names given, put the query csv file in ratings.csv, and run parse.py


2016 Results

Jelly:                1906
Pyth:                 1711
GS:                   1674
Japt:                 1653
MATL:                 1607
Dyalog APL:           1580
J:                    1579
CJam:                 1578
Jolf:                 1559
Seriously:            1535
TeaScript:            1499
GolfScript:           1497
Pip:                  1478
Burlesque:            1460
APL:                  1443
O:                    1427
Vitsy:                1424
𝔼𝕊𝕄𝕚𝕟:                1384
Vim:                  1371
q:                    1355
K:                    1322
TI:                   1301
sh:                   1294
Octave:               1289
PARI/GP:              1272
dc:                   1269
GNU sed:              1229
Retina:               1213
x:                    1209
Bash + coreutils:     1207
Bash + GNU utilities: 1201
bc:                   1189
golflua:              1187
Sed:                  1182
HTML:                 1168
Julia:                1165
QBasic:               1156
Sage:                 1146
Ruby:                 1124
AWK:                  1101
ZSH:                  1086
Mathematica:          1083
Gema:                 1079
Bash:                 1079
jq:                   1060
Haskell:              1051
MATLAB:               1042
PowerShell v:         1028
Perl:                 1028
BBC BASIC:            1022
Piet:                 1019
Score:                1018
Shell script:         1017
Mumps:                1015
Languages:            1013
ECMAScript:           1012
Regex:                1000
LiveScript:           993
Tcl:                  989
CoffeeScript:         988
Groovy:               986
Postscript:           984
Windows PowerShell:   981
Powershell:           966
R:                    964
Befunge:              964
gawk:                 951
AutoHotkey:           934
Javascript ES:        931
Smalltalk:            928
STATA:                926
Matlab/Octave:        926
REBOL:                923
SAS:                  917
Clojure:              913
ES:                   909
PHP:                  908
Dart:                 908
Python:               894
JavaScript:           894
Node.js:              893
F#:                   880
Marbelous:            875
Excel VBA:            858
AutoIt:               853
Shell:                850
Common Lisp:          845
Brainfuck:            843
Lua:                  835
Processing:           832
Scheme:               824
CSS:                  823
Pure bash:            821
SWI:                  786
Game Maker Language:  785
FORTH:                778
SpecBAS:              776
Scala:                772
Cobra:                769
HTML + JavaScript:    758
Whitespace:           750
Applescript:          739
T:                    724
Batch:                720
Fortran:              715
Prolog:               711
Turing Machine Code:  708
C:                    706
SQL:                  704
VBA:                  703
Factor:               700
Kotlin:               690
JS:                   686
Pascal:               685
Erlang:               674
Emacs Lisp:           673
OCaml:                668
Racket:               658
VBScript:             654
VB.NET:               643
Swift:                636
D:                    621
Windows Batch:        612
Ceylon:               591
C++:                  587
Java:                 559
Oracle SQL:           548
Delphi:               539
Go:                   537
C#:                   535
Rust:                 528

2020 Results

Pyke:                 1777
Jelly:                1678
Pyth:                 1675
Japt:                 1665
MATL:                 1604
Jolf:                 1602
Actually:             1583
CJam:                 1545
gs:                   1506
K:                    1497
Dyalog APL:           1488
Seriously:            1472
Convex:               1470
GolfScript:           1448
J:                    1446
Pip:                  1442
𝔼𝕊𝕄𝕚𝕟:                1432
TeaScript:            1421
Vitsy:                1408
v:                    1344
Burlesque:            1317
q/kdb+:               1301
Retina:               1281
TI:                   1276
SED:                  1239
Bash + GNU utilities: 1237
ZX Spectrum BASIC:    1236
Q:                    1224
APL+WIN:              1224
Gema:                 1221
PARI/GP:              1220
Vim:                  1207
x:                    1183
MATLAB / Octave:      1171
APL:                  1165
Octave:               1163
Perl:                 1151
Cheddar:              1130
jq:                   1116
Wolfram Language:     1115
Ruby:                 1113
Mathematica:          1112
flag) =:              1095
BBC Basic:            1094
Bash + coreutils:     1093
Casio Basic:          1089
JavaScript:           1085
GNU sed:              1066
Windows PowerShell:   1064
Haskell:              1061
Julia:                1049
ECMAScript:           1046
Pure Bash:            1039
SmileBASIC:           1029
Smalltalk:            1029
Maple:                1026
Autohotkey:           1018
Rebol:                1010
R:                    1006
CoffeeScript:         993
Zsh:                  991
Fourier:              990
AWK:                  989
dc:                   986
Casio:                982
Powershell:           980
PowerShell v:         980
Matlab:               976
Mumps:                974
Bash:                 955
Excel:                953
Postscript:           950
Python:               948
Google Sheets:        929
Befunge:              929
ES:                   918
HTML:                 905
VBA:                  902
REXX:                 901
Javascript ES:        897
Tcl/Tk:               882
FORTH:                876
Groovy:               867
JS:                   846
PHP:                  842
Cobra:                840
Excel VBA:            838
Dart:                 837
Axiom:                834
Scala:                830
Node.js:              826
HTML + JavaScript:    826
QBasic:               825
CSS:                  823
Java:                 812
Tcl:                  805
Brainfuck:            802
Processing:           800
Clojure:              781
AppleScript:          776
Scheme:               771
Erlang:               768
SWI:                  753
Emacs Lisp:           750
Factor:               747
C:                    738
Delphi:               728
VBScript:             720
Swift:                715
Kotlin:               712
Lua:                  711
Whitespace:           708
TSQL:                 702
SpecBAS:              701
Prolog:               699
Common Lisp:          695
F#:                   673
Ceylon:               664
Rust:                 663
SAS:                  657
T:                    652
Batch:                639
Windows Batch:        636
GameMaker Language:   632
VB.NET:               631
SQL:                  623
Pascal:               620
C#:                   619
D:                    604
Oracle SQL:           556
Racket:               550
C++:                  516
Fortran:              508
Go:                   467



SCALE = 400
K = 10
PAD = 50

import itertools
def concat(x):
    return list(itertools.chain.from_iterable(x))

class Ratings:
    def __init__(self, challenges):
        langs = concat(challenges)
        self.ratings = {lang: STARTING for lang in langs}
        for challenge in challenges:

    def update_game(self, first, second, score):
        f_rat = self.ratings[first]
        s_rat = self.ratings[second]

        f_exp = 1/(1 + 10 ** ((s_rat - f_rat) / 400))
        s_exp = 1/(1 + 10 ** ((f_rat - s_rat) / 400))

        f_upd = K * (score - f_exp)
        s_upd = K * ((1 - score) - s_exp)

        self.ratings[first] += f_upd
        self.ratings[second] += s_upd

    def update_challenge(self, scores):
        for lang1 in sorted(scores):
            for lang2 in sorted(scores):
                if lang1 > lang2:
                    s1 = scores[lang1]
                    s2 = scores[lang2]
                    result = 1 if s1 < s2 else 0.5 if s1 == s2 else 0
                    self.update_game(lang1, lang2, result)

    def __str__(self):
        return '\n'.join(key + ':' +
                ' ' * (max(len(lang) + 1 for lang in self.ratings) - len(key)) + str(val)
                         for key, val in sorted(((lang, int(self.ratings[lang]))
                             for lang in self.ratings),
                             key=lambda x:x[::-1], reverse=True))


import csv
import re
import elo


lines = []
scores = {}
challenges = []
lang_counts = {}
canon_to_actual = {}

with open('ratings.csv') as csvfile:
    reader = csv.DictReader(csvfile)
    for row in reader:
        lines.append((row['Body'], row['ParentId']))

for line, challenge in lines:
    line = re.sub("<s(trike)?>.*</s(trike)?>", "", line)
    match = re.match("<.*>(.*)</.*>", line)
    if match:
        line = match.group(1)
        lang = re.match("([^-,\(\d:]*)", line)
        if lang:
            lang = lang.group(1).strip()
            canon = lang.lower().replace(' ', '')
            canon_to_actual[canon] = lang
            score = re.search("(\d+)\s*\w*\)?$", line)
            if score:
                score = int(score.group(1))
                c = scores.get(challenge, [])
                c.append((canon, score))
                scores[challenge] = c
                lang_counts[canon] = lang_counts.get(canon, 0) + 1

for c in sorted(scores, key=int):
    results = scores[c]
    c_map = {}
    for canon, score in results:
        if lang_counts[canon] > CUTOFF and re.match("\w", canon):
            lang = canon_to_actual[canon]
            old_score = c_map.get(lang, score)
            c_map[lang] = min(score, old_score)

  • 2
    \$\begingroup\$ Great list, and it also represent well what's the reality. Actually, there's one more thing to do: make it case insensitive. There's for instance, one entry for Javascript ES and an other one for JavaScript ES \$\endgroup\$ – Katenkyo Apr 5 '16 at 11:23
  • \$\begingroup\$ This seems to get around the problem of some languages only being used on easy challenges, since two languages are only compared on challenges they have in common. \$\endgroup\$ – trichoplax Apr 5 '16 at 13:58
  • \$\begingroup\$ I also like that a language is scored by its shortest answer to a question, so it approximates the best a language can do rather than the average performance across a number of different users. \$\endgroup\$ – trichoplax Apr 5 '16 at 14:00
  • \$\begingroup\$ @Katenkyo Fixed the capitalization problems. \$\endgroup\$ – isaacg Apr 5 '16 at 14:56
  • \$\begingroup\$ The Elo system might actually be better that you say, assuming you parse the answers in posted-order. Languages here evolve over time. \$\endgroup\$ – Nathan Merrill Apr 5 '16 at 18:23
  • \$\begingroup\$ @NathanMerrill They're sorted by question age, so you might be right. \$\endgroup\$ – isaacg Apr 5 '16 at 19:30
  • 2
    \$\begingroup\$ I think one of the reasons jelly ranks so high is that dennis is the only one who knows how to use it. =) \$\endgroup\$ – James Apr 22 '16 at 4:14
  • \$\begingroup\$ @DJMcMayhem: I don't think so. I think Dennis uses Jelly because it's short (as do I), rather than Jelly being short because Dennis uses it. (You can observe that Dennis tends to do equally well, relative to other entries in the same language, when using a language other than Jelly; but the entry ends up overall longer.) \$\endgroup\$ – user62131 Apr 25 '17 at 16:13
  • \$\begingroup\$ @James I think one of the reasons pyke ranks so high is that blue is the only one who knows how to use it. =) \$\endgroup\$ – petStorm Feb 17 at 4:07

K-means clustering by code length

Note: This is just a first-pass attempt with a very simple method. There are certainly better ways to approach this but at least this is a start.

My approach here was to get all answers to code golf challenges without bonuses that had (at least somewhat) appropriately formatted post headers, parse out the language name and byte count from the header, perform k-means clustering on the sizes with 6 groups, and for each language select the cluster into which most of the sizes for that language fell. Languages which have been used fewer than 10 times on the site, as well as any answer with a byte count greater than 170, were excluded.

The clusters are listed in ascending order of the values of their cluster centers.

Group 1, mean 12.03099

APL            GS2             Rebmu
Arcyou         Hexagony        rs
AutoIt         J               Sed
Batch          Japt            Self
BC             Jelly           Seriously
Beeswax        Jolf            Stuck
Befunge        K               TeaScript
Brachylog      Labyrinth       TI-BASIC
Brainfuck      Marbelous       Vim
Burlesque      MATL            Vitsy
Candy          Microscript     zsh
Clip           Minkolang
DC             Mouse
DUP            O
Element        Piet
ESMin          Pip
Factor         Platypar
Fission        Prelude
Gema           Pyth
GolfScript     QBASIC

Group 2, mean 31.75217

Milky Way

Group 3, mean 54.14769

Game Maker Language

Group 4, mean 79.65123

BBC Basic        Scala
CoffeeScript     SQL
Common Lisp      Swift
CSS              TCL
Emacs Lisp       VB.NET
Fortran          VBScript

Group 5, mean 109.9524


Group 6, mean 147.5168


Thoughts on this grouping

This is roughly what one might expect, with the golfing languages in tier 1 and Java in tier 5. This particular grouping appears to be somewhat of a success in that regard. However, 6 was probably too many groups; 4 or 5 would likely be sufficient. Heirarchical clustering may be another interesting route for analysis since no number of classes is prespecified.


This should not, under any circumstances, be used to impose any kind of language handicap or restriction. It's merely for the sake of interest and giving folks a group to mentally compare their scores to, should they feel the need.

For the curious

The data were collected via this Stack Exchange Data Explorer query. The analysis was done in R and is available via this Gist.

  • 14
    \$\begingroup\$ I can spot one fatal flaw in this approach: challenges that would require significantly long code in golfing languages are usually not answered with those languages. For example, as the source of almost all of the Seriously answers, I admit that I usually don't try to answer a challenge with Seriously unless it's under 40 or so bytes (because any longer becomes unmanageable thanks to my poor design decisions). This skews Seriously towards a higher tier than what it probably should be. \$\endgroup\$ – Mego Mar 27 '16 at 7:00
  • 1
    \$\begingroup\$ @Mego Yep. There's a whole lot I didn't take into account. This is just a means of getting the ball rolling on an approach in the simplest possible way. \$\endgroup\$ – Alex A. Mar 27 '16 at 7:04
  • \$\begingroup\$ 404 on the gist. When you say "perform k-means clustering on the sizes with 6 groups", what exactly was the input to the clustering algorithm? \$\endgroup\$ – Peter Taylor Mar 27 '16 at 17:56
  • \$\begingroup\$ @PeterTaylor Looks like I accidentally deleted the gist. I've readded it and updated the URL. The input was the code lengths as parsed from the headers. Those were grouped into 6 classes using k-means and merged back to their respective languages. Then, for each language, I selected the class into which it fell most often. \$\endgroup\$ – Alex A. Mar 27 '16 at 18:08
  • \$\begingroup\$ As an avid vim user and a vim golfer, I'm really happy to see it in the same tier as the golfing languages. =D \$\endgroup\$ – James Mar 28 '16 at 1:34
  • 7
    \$\begingroup\$ I was quite surprised to see QBASIC in the top tier, until I looked at the QBASIC answers on the site. It's pretty good backup for Mego's comment. I suppose there's probably a somewhat better way to do this, but I'm not really sure it won't be filled with these outliers as well. \$\endgroup\$ – Geobits Mar 28 '16 at 2:23
  • \$\begingroup\$ Perhaps sampling from a set of challenges with many answers in many different languages would be better, rather than across the entire site, to avoid the problem I brought up. Granted, I still don't think this will result in anything useful. \$\endgroup\$ – Mego Mar 28 '16 at 7:04
  • 2
    \$\begingroup\$ I think that @Mego's comment deserves some thought. Perhaps we should score based on relation to the median answer of each question? \$\endgroup\$ – Not that Charles Mar 28 '16 at 14:31
  • 3
    \$\begingroup\$ What poor souls answered in Delphi? \$\endgroup\$ – Morgan Thrapp Mar 28 '16 at 18:34
  • 1
    \$\begingroup\$ I think BATCH is in the first group because not many people answer in BATCH, and, when they do, it's usually on an easy question. \$\endgroup\$ – Conor O'Brien Apr 3 '16 at 22:36
  • \$\begingroup\$ @CᴏɴᴏʀO'Bʀɪᴇɴ more reasons why this is biased -- cool, but biased \$\endgroup\$ – cat Apr 5 '16 at 1:21
  • \$\begingroup\$ @Mego this is probably how Brainfuck ended up in group 1 \$\endgroup\$ – Felix Dombek Oct 6 '17 at 2:25

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .