Houdini 2.0c x64’s top choices were:
My main concern with both engine’s first choice of 13.Qxa8 is that after 13…Nc6 the White Q is trapped and will have to be sacrificed for 2R’s which is evaluated nearly equal, but I am always very skeptical of engine evaluations in unbalanced material situations because they tend to put a lot of emphasis on material. As for 13.Nh4, it was played in Gorovykh (2409)-Jedlicka (2265), Pardubice 2007 and Black won though there were some improvements in White’s play that were worth considering. It’s also interesting that in that game White also did not give up his Q for 2R’s when the possibility presented itself.
Houdini’s 13.a4, isolating Black’s b-Pawn may be worth considering also, but it simply does not appeal to me and 13.Nh4 looks too risky. So, that leaves 13.Be3 which my gut feeling says is best because, if nothing else, it develops a piece. This seems to be a good place to try the MCA to see what the trends of each move might be so I plugged in Deep Rybka 4 and let it play about a thousand games with each move with the following results:
13.Bd2 – Black wins 3:1
13.Qxa8 - Black wins 2:1
13.a5 – This resulted in about a 50/50 result
13.Nh4 - Black wins 2:1
13.Be3 - White scores +21% -29% =50%
Basically the MCA results will inform you how drawish a position is or whether one side is winning but it's not good at telling you what move is the actually the best. So, in the end, even though 13.a5 scored the best for White, I selected 13.Be3 because it’s a developing move and deeper analysis convinced me it offers White better chances than 13.a5.
I think this feature, if you have access to it, could be a handy analysis tool in those situations where you need a second opinion, but I would not rely entirely on the moves it produces as being the best.