When do people usually resign? Some statistics

collected data from 236 users who played live at the moment on OGS and had at least one 19x19 even game which they resigned against human in their history
1 even 19x19 game from each user, newest one which they resigned

it includes those who barely know how to play and those who resigned for weird reasons
so, people estimate surprisingly good

if we exclude incorrect resigns:
168,19 is average move number at which people usually resign
people are behind by 63,17 points on average when they resign. Estimated by KataGo

ranks are from the moment of resign, not current rank
raw data of ranks sorted by rank:

vertical more than 0 are dans, below 0 are kyu.
4th dan highest, below 25k usually displayed as 25k but can be seen in game history
there are also 6 [?] users
horizontal: just a number of point

raw data of move number at which users resign, sorted by number of move

vertical: move number
horizontal: just a number of point

raw data of score by which they were behind when resigned, sorted by score (vertical)

below 0 means they were ahead when resigned
horizontal: just a number of point

ranks histogram:

vertical: number of users who are similar enough to rank written on horizontal axis
smoothing is Â±1 rank. So 6k, 5k and 4k are similar enough to 5k, so height of 5k is sum of 6k, 5k and 4k
most are sdk

move at which they resign histogram:

vertical: number of resigns on move which is similar enough to move number written on horizontal axis
smoothing is Â±10. So moves 160 and 180 are similar enough to move 170. Height of 170 column is sum of resigns from move 160 to move 180

score histogram:

horizontal axis: by how many points behind they were when resigned. Below 0 means they were ahead
vertical axis: number of resigns with score difference similar enough to written on horizontal axis
smoothing is Â±10. So 30 and 50 are similar enough to 40. Column of â€ś40 points behindâ€ť has height of sum of number of resigns from score difference 30 to score difference 50

raw data where horizontal axis is rank (below 0 are kyu, above 0 are dans)
vertical axis: score difference. Below 0 means resign when ahead

raw data where horizontal axis is rank (below 0 are kyu, above 0 are dans)
vertical axis: move at which they resign

raw data where horizontal axis is move at which they resign
vertical axis: score difference. Below 0 means resign when ahead

for more graphs look here:
https://forums.online-go.com/t/when-do-people-usually-resign-some-statistics/52977/13?u=stone.defender

11 Likes

Itâ€™s nice to have confirmation of my guesses. I was just surprised by the 7% of winning resigns, a bit high to my taste.

1 Like

• Someone called them, or they have an appointment, or other real-life reasons.
• They got angry after one of their groups got killed, or after making a big mistake.
• They were ahead according to AI, but the AI found a tesuji/a way to live/a way to kill that they couldnâ€™t see. Other humans of similar level would also judge they were behind.
• Sandbagging.
• They got bored by their opponentâ€™s stalling moves.
3 Likes

I wonder if we could build some similar analysis from the big bulk download of millions games.

One information, the coolest one, would be missing for sure: the score estimation at the moment of resigning. But probably we have some of the others (number of moves, maybe also ranksâ€¦ I canâ€™t rememberâ€¦ ) and we could be able to have same analysis on other boards.

• They miscounted
• They accidentally resigned (mis-click on the resign button and the confirmation)
2 Likes

Great data.

I was told: â€śnever resign before move 150â€ť when I had a teaching game in early DDK.

Are the histograms normalised by population at that rank?

1 Like

which? how? why?

1 Like

This topic speaks to me, because I resigned while I was ahead todayâ€¦ Because it was a correspondence game and I forgot what color I was, and saw that black was heavily behind.

Doh!

Anyway, really nice statistical analysis here.

What does your data pipeline look like here? Do you go straight from game SGF into katago and then shoot it to a spreadsheet, or are you doing something more complex (and interesting)?

I love performing this kind of analysis, and Iâ€™m new to go and analyzing its data in general, so whatever level of detail you want to go into is fine by me!

10 Likes

I meant â€śit appears that the histograms by rank have raw resignation countâ€ť â€¦ so if there is a rank band with only 2 games in it, it only shows as â€ś2 resignationsâ€ť, but thatâ€™s 100% of the population of eligible games.

As I type that I can see that Iâ€™m not clear either on â€śhow would this be normalizedâ€ť, but it does seem odd?

Love it!

Even @yebellz didnâ€™t think of this one

4 Likes

Thatâ€™s good in early ddk

Itâ€™s good later to progressively learn to evaluate the board and resign. The stronger the quicker becomes possible because you know that you donâ€™t stand a chance to come back playing someone strong like you.

[?] ranked are excluded, resigns when ahead are excluded
all graphs are average 80, so for example the most left dot of next graph is average rank of 80 weakest users, the most right dot of next graph is average rank of 80 strongest users
look at 5k dot, if you look at the same horizontal position on graph below, it has value 175, it means that users who are average 5k usually resign at move average 175
these 9 graphs may be useful to see some correlations

All data is sorted by rank
average rank:

use this graph as horizontal axis for next two graphs
average move: (of resign)

average score: (by how much points behind when resigned)

All data is sorted by move
average move: (of resign)

use this graph as horizontal axis for next two graphs
average score: (by how much points behind when resigned)

average rank:

All data is sorted by score
average score: (by how much points behind when resigned)

use this graph as horizontal axis for next two graphs
average move: (of resign)

average rank:

1 Like

or not
probably not enough

for different rank different number of points to average should be used depending on how many people are in this rank. And ranks with not enough data should be painted less clear.