Sunday, June 7, 2020

Analysis - Wait Win Rates in Riichi Hands

I ran through 1.2 million Houou replays and checked the wait of every riichi hand, as well as whether it won or was furiten. Let's see the results!

Now, there are a LOT of different possible waits for a hand to have in riichi mahjong. There are 395 unique waits in this dataset, and many more shapes, so we'll only look at a small subset of them.

People often refer to different classes of ryanmen. A 1-4 or 6-9 ryanmen is a class A, while a 3-6 or 4-7 is a class C. The idea being that the further outside the tiles are, the more likely they'll be cut. Let's take a look at a section of the data showing some common waits.

The "Width" column lists the maximum width for that wait. The "Total" column is the overall winrate for all instances of that wait. The numbered columns are how many outs are visible on the board and in the riichi player's hand.
As we can see from the second section here, a class A ryanmen has a notably higher winrate than a class C. A class A ryanmen with two of its outs visible has a win rate comparable to a class C ryanmen with none visible.

People also often say that a shanpon involving honors has similar winrates to a ryanmen, and we can see that here. Honor shanpons are very strong with no outs, and with one out they're similar to ryanmen with two out. A double honor shanpon with no out is even comparable to a sanmenchan!

Speaking of sanmenchan, I've often heard it said that a furiten sanmenchan is about as good as a non-furiten ryanmen. Let's take a look at the furiten data.
Well, it does seem to match up to ryanmen with the equivalent number of outs. A furiten sanmenchan with one out visible has a winrate a bit lower than a ryanmen with one visible, and so on.

That's a taste of what the data has in store. Make sure to check out the full spreadsheet, where I've gathered eighteen common shapes in a pretty format. It also includes example hands for every wait + outs combination and the percentage of wins that are tsumo for each wait.

This analysis took like twenty hours to run and I ended up doing it four or five times.

1 comment:

  1. This data helped me settle an internet dispute, 10/10 thanks for all the hard work

    ReplyDelete