8

To specify the question a bit:

  1. I consider a blunder a move which worsens the position by 3 pawns (300 centipawns) or more. I think this is a pretty good base mark, since a strong computer will generally judge sacrificial moves accurately.
  2. Lets only consider amateur (<2100 rating) games, since beyond this, such blunders are very rare

A breakdown by rating would be very nice

EDIT: In consideration of the numerous comments I've decided to tighten the bolts some more on this question.

Addendum 1: The question is asking how many games out of a pool of games will contain a blunder. Not how many moves within a game are blunders.

Addendum 2: It seems 100 centipawn shifts are common even in master level games, so I have upped the ante to 300 centipawns, these would be blunders that outright lose the game.

Addendum 3: I am not asking for how many games are won or lost through a blunder (this is basically impossible to determine), but just how many contain at least one.

Dider
  • 841
  • 9
  • 14
  • Just a little thought experiment: suppose players A and B are playing optimally (perfectly) for 20 moves, then A hangs his queen and should be dead lost, but B misses it entirely so blunders back to an even position, and the game ends in a draw after another 20 optimal moves. Shall we say the outcome of that game was decided a blunder, or no? ... (cont'd) – ETD Mar 10 '15 at 00:41
  • (cont'd) ... Maybe we say the outcome was decided by the combination of blunders that occurred in the game; that'd be fine, but in a sense that's always true, even when there are none. (I guess my point is that I think it'd be hard to definitively answer anything beyond the concrete question of what % of games feature a blunder, with "blunder" defined relative to whatever fixed particular engine evaluation you like.) – ETD Mar 10 '15 at 00:43
  • 1
    Good point, I changed the title to clarify – Dider Mar 10 '15 at 00:43
  • 2
    " (<2100 rating) games, since beyond this, such blunders are very rare" I vehemently disagree. – Cleveland Mar 10 '15 at 01:49
  • Really? I'm not within that rating range so I wouldn't know, but I always thought (from looking at master games) that blunders (as I defined them) are quite rare at the master level – Dider Mar 10 '15 at 01:53
  • 3
    https://medium.com/pachyderm-data/when-grandmasters-blunder-a819860b883d At the 2200 level, about 4% of moves are blunders of >1.00. That's one every twenty half-moves, or four such blunders in a 40-move game. – Cleveland Mar 10 '15 at 02:03
  • 3
    Below 2100, I would say that 99% of games contain a blunder. – Tony Ennis Mar 10 '15 at 04:59
  • 1
    Note that the definitions of blunder in this question and in the article Cleveland links to share a flaw in that they consider a move that lowers the evaluation from +6 to +5 a blunder, even though the move played is still clearly winning. I assume that a noticeable number of "blunders" in that article's dataset fall into that sort of category. A better definition would use expected value rather than evaluation. – dfan Mar 10 '15 at 13:39

1 Answers1

10

I happen to have a dataset with 25000 games with stockfish evaluations after every move. I just did a little blunder search and these are my findings:

Blunders from an equal position (-1.00 < eval < 1.00) are very rare, even among weaker players. That is not particularly surprising, because we tend to leave the equality region in little steps during the opening and the blunders come when we are under real pressure and low on time.

  • Elo: 1500: 100cp Blunder every 45 moves.
  • Elo: 1600: 100cp Blunder every 46 moves.
  • Elo: 1700: 100cp Blunder every 50 moves.
  • Elo: 1800: 100cp Blunder every 52 moves.
  • Elo: 1900: 100cp Blunder every 55 moves.
  • Elo: 2000: 100cp Blunder every 59 moves.
  • Elo: 2100: 100cp Blunder every 64 moves.
  • Elo: 2200: 100cp Blunder every 69 moves.
  • Elo: 2300: 100cp Blunder every 74 moves.
  • Elo: 2400: 100cp Blunder every 78 moves.
  • Elo: 2500: 100cp Blunder every 82 moves.
  • Elo: 2600: 100cp Blunder every 85 moves.
  • Elo: 2700: 100cp Blunder every 85 moves.

Really big blunders >300cp from an equal position are even more rare. Which also isn't particularly surprising, after all even if you blunder a full piece you usually get a little something in return.

  • Elo: 1500: 300cp Blunder every 223 moves.
  • Elo: 1600: 300cp Blunder every 241 moves.
  • Elo: 1700: 300cp Blunder every 274 moves.
  • Elo: 1800: 300cp Blunder every 299 moves.
  • Elo: 1900: 300cp Blunder every 335 moves.
  • Elo: 2000: 300cp Blunder every 352 moves.
  • Elo: 2100: 300cp Blunder every 381 moves.
  • Elo: 2200: 300cp Blunder every 421 moves.
  • Elo: 2300: 300cp Blunder every 446 moves.
  • Elo: 2400: 300cp Blunder every 479 moves.
  • Elo: 2500: 300cp Blunder every 509 moves.
  • Elo: 2600: 300cp Blunder every 525 moves.
  • Elo: 2700: 300cp Blunder every 530 moves.

If you leave out the requirement of an equal position, the numbers go way up. As opposed to the relevancy ….

  • Elo: 1500: 100cp Blunder every 9 moves.
  • Elo: 1600: 100cp Blunder every 9 moves.
  • Elo: 1700: 100cp Blunder every 10 moves.
  • Elo: 1800: 100cp Blunder every 10 moves.
  • Elo: 1900: 100cp Blunder every 11 moves.
  • Elo: 2000: 100cp Blunder every 12 moves.
  • Elo: 2100: 100cp Blunder every 13 moves.
  • Elo: 2200: 100cp Blunder every 14 moves.
  • Elo: 2300: 100cp Blunder every 14 moves.
  • Elo: 2400: 100cp Blunder every 15 moves.
  • Elo: 2500: 100cp Blunder every 16 moves.
  • Elo: 2600: 100cp Blunder every 16 moves.
  • Elo: 2700: 100cp Blunder every 17 moves.

And the bigger blunders:

  • Elo: 1500: 300cp Blunder every 23 moves.
  • Elo: 1600: 300cp Blunder every 25 moves.
  • Elo: 1700: 300cp Blunder every 27 moves.
  • Elo: 1800: 300cp Blunder every 29 moves.
  • Elo: 1900: 300cp Blunder every 32 moves.
  • Elo: 2000: 300cp Blunder every 34 moves.
  • Elo: 2100: 300cp Blunder every 37 moves.
  • Elo: 2200: 300cp Blunder every 40 moves.
  • Elo: 2300: 300cp Blunder every 43 moves.
  • Elo: 2400: 300cp Blunder every 46 moves.
  • Elo: 2500: 300cp Blunder every 48 moves.
  • Elo: 2600: 300cp Blunder every 49 moves.
  • Elo: 2700: 300cp Blunder every 50 moves.

Those numbers are a little higher than in the linked blog, that's probably because stockfish likes to throw out big numbers compared to other engines.

To confer to the addendum and the final phrasing of the question (i.e. percentage of games with a 300cp blunder.):

  • Elo: 1500: 62 percent of the games contain a blunder.
  • Elo: 1600: 58 percent of the games contain a blunder.
  • Elo: 1700: 53 percent of the games contain a blunder.
  • Elo: 1800: 50 percent of the games contain a blunder.
  • Elo: 1900: 47 percent of the games contain a blunder.
  • Elo: 2000: 45 percent of the games contain a blunder.
  • Elo: 2100: 41 percent of the games contain a blunder.
  • Elo: 2200: 37 percent of the games contain a blunder.
  • Elo: 2300: 33 percent of the games contain a blunder.
  • Elo: 2400: 29 percent of the games contain a blunder.
  • Elo: 2500: 25 percent of the games contain a blunder.
  • Elo: 2600: 24 percent of the games contain a blunder.
  • Elo: 2700: 24 percent of the games contain a blunder.

With the requirement of an equal position:

  • Elo: 1500: 13 percent of the games contain a blunder.
  • Elo: 1600: 12 percent of the games contain a blunder.
  • Elo: 1700: 10 percent of the games contain a blunder.
  • Elo: 1800: 10 percent of the games contain a blunder.
  • Elo: 1900: 8 percent of the games contain a blunder.
  • Elo: 2000: 8 percent of the games contain a blunder.
  • Elo: 2100: 7 percent of the games contain a blunder.
  • Elo: 2200: 5 percent of the games contain a blunder.
  • Elo: 2300: 6 percent of the games contain a blunder.
  • Elo: 2400: 4 percent of the games contain a blunder.
  • Elo: 2500: 3 percent of the games contain a blunder.
  • Elo: 2600: 3 percent of the games contain a blunder.
  • Elo: 2700: 3 percent of the games contain a blunder.
BlindKungFuMaster
  • 19,216
  • 1
  • 50
  • 79
  • Could you specify what is considered an "equal position"? Is it an evaluation of exactly 0.0 or is there is an interval? – Dider Jun 23 '15 at 17:16
  • At the beginning of my answer I defined equal as (-1.00 < eval < 1.00). Which is quite a margin, but the essential thing wasn't so much the equalilty, but the relation between the magnitude of the blunder and the preexisting disadvantage/advantage. – BlindKungFuMaster Jun 24 '15 at 08:05