Statistically correct moves and openings

sanath1987: Thanks for your interesting insights!

Differentiation2: Thanks a lot! But U forgot that we allow 0.1 margin. Since ev(1. Nf3) =ev(1. c4)=+0.5, we should consider that statistically correct replies have evaluations =< 0.6. Even Anglo-Indian can be round down to 0.6 and considered as statistically correct. Maybe we should switch to chesstempo's database (2200+ vs 2200+) where there are more accurate pourcentages:

https://chesstempo.com/game-database.html

Using this database, the evaluation of Anglo-Indian is 0.613. Do U agree with these propositions?

I'm working on evaluations of replies to 1. g3 using the chesstempo's database.

As far as Anti-Benoni and 4-knights comparison goes, the four-knights seem to retain the symmetry longer than Anti-Benoni. In the c4 openings, there are a lot of pawn breaks available for both sides which mostly breaks the symmetry sooner or later.

As far as Sicilian and French better than KID or Pirc, we might have to look a bit deeper in mainlines of major variations and we might well be surprised with the evaluation. What is the statistics between Advance French vs Classical French and Open Sicilian vs Closed Sicilian? That might shed some light for better understanding of the opening ideas.

All right.

I think strong grandmasters often choose irregular opening moves when facing off against players who are weaker than themselves because they know they can win in the middlegame or ending thanks to superior technique.  If you look at just win draw loss statistics, this makes it look as if poor moves are strong, even though the result of each game has little to do with the opening chosen.

When strong grandmasters face players higher rated than themselves, they are much more likely to choose moves that they consider theoretically strongest, because they don't want to end up in a worse position coming out of the opening.

Another factor is in rare lines, players get out of book rather early in the game, and both may make a long series of mistakes in the opening.  What you really want to be looking at are games where both sides are playing good moves, otherwise it's hard to know what to make of the results.

There are a couple of ways to get around these problems.  In the first case, you can look at the performance Elo achieved by a certain opening move in an archive of games played by strong grandmasters.

In the second case, you could backsolve the games by result, eliminating lines where one player seems to have made a mistake early on.  You also want to eliminate games where the result seems to have little to do with the opening.  This seems to be what GMs like Boris Avrukh do when they are preparing analysis for publication.

If you are looking to take this idea further, the next step could be to incorporate the ratings of the players involved.

When I started to explore chess openings seriously, I noticed this actually had a big influence on how various openings scored. 

Some openings are often chosen by strong players playing for a win with black against weaker players, causing these black openings to look better than they are. 

Also some openings are often chosen by weaker white players who are content with a draw. For example, in my database, 1. e4 e5 2. Nf3 Nc6 3. Bb5 Nf6 4. O-O Nxe4 5. Re1 gives black a plus score.  But it is probably impossible to argue that black is better in this line - white gets a symmetrical position simply up a couple of tempi.  Why is this happening?  Well, the positions are drawish, and so lower rated white players are often content to play them - indeed the average white player in this position is lower rated than the average black player, and so black still ends up winning a little more than white.

Whoa, wayne_thomas and I both posted something a bit similar less than a minute apart

Statistical correct replies to Benko's opening 1. g3 (ev=+0.42):

• 1...d5 Benko: QP, ev=+0.30
• 1...e5 Benko: KP, ev=+0.16 statistically excellent
• 1...c5 Benko: Sicilian, ev=+0.32

Statistically incorrect replies:

• 1...g6 Benko: Symmetrical, ev=+0.55
• 1...Nf6 Benko: Indian, ev=+0.88
• 1...f5 Benko: Dutch, ev=+0.59

Let's make a break for today and think a little. Maybe we should also discuss/modify 0.1 margin since chesstempo's database allows to calculate up to 2 decimal digits ? It's also necessary to discuss what to do with border cases like Anglo-Indian. Any ideas?

wayne_thomas & NM happytoad: Thanks for your insights and useful comments! If we remain motivated enough, we'll refine our analysis.

All right. I have an idea. Let's get rid of the margin but introduce instead 2 subcategories: statistically good and statistically suboptimal. A statistically good move m: X->Y is when ev(Y)>=ev(X) when white moves (and vice versa when black moves). Now, if ev(Y)<ev(X) but close to it, we'll say that it's statistically suboptimal, otherwise statistically incorrect. Now, we should only quantify this notion of close evaluations. I'd propose 0.2, in this way Anglo-Indian as well as Benko: Symmetrical and Benko: Dutch will go into suboptimal category while Benko: Indian will remain in the statistically incorrect category. What do U think about it?

Great!

Tomorrow I'll modify the first post making the summary of our calculations so far using the chesstempo's database and modified definitions. Then we can discuss it and go further.

Personally, I'm working on my French Defense as a response to 1.e4.  I see it has a "correct response" evaluation which is nice.  I like it because it feels solid and it seems that white needs more theory than black for correct play.

However, I've noticed a tendency for the advanced variation to leave book early in my games.  1.e4  e6  2.d4  d5 3.e5  c5  4.c3 or Nf3  Nc6 5.Bb5   This isn't taken as a serious move in any books I've looked at and so I'm thinking it is simply a mistake for white.  I usually respond 5 ... Qb6 6.Bxc6  bxc6.  Now I have to deal with a backward pawn, my Knight is no longer exerting pressure on the center. I look forward to c5 again but that seems to be easily ignored after c3 or Nf3 and I've got King side development concerns.  I know that 5Bb5 is supposed to be statistically flawed, I just struggle to clearly see the path to proving that in game.

ArcBlade: The French defense is statistically correct but on the edge, far behind Sicilian. In our new classification, it will become statistically suboptimal. In your first Advance Variation line, 4. c3 is statistically good, it has also good engine evaluations. In your second line, 5. Bb5 can't be measured statistically, there are not enough master games. But this move is not a mistake or even inaccuracy, it's just suboptimal. It's not easy to prove it. Imho, the best black's replies on 5. Bb5 are 5...cxd4 and 5...Bxd7.

Interesting stuff.  The Modern Defense's strength surprises me. For 1700 and under players who have played it against me, I simply point my pieces at my opponent's King and prepare to attack him while Black makes all those passive, non-threatening opening moves.

wayne_thomas:  That was how I played it in my most recent game.

Why 11...Qb6 moving your queen from a central square to a less central one?  It wasn't even attack yet.  How about 11...Bb4 threatening to win a second pawn?

Incidentally, you emerged from the opening a pawn to the good with an incredibly solid pawn centre.  Black must be winning.