Statistically correct elo estimates from SPRT tests

[Michel]

It is well known that naive elo estimates after a finished SPRT test (like e.g. those returned by cutechess-cli) are quite biased. The following web page computes statistically correct elo estimates for finished SPRT tests.

http://hardy.uhasselt.be/Toga/sprt_elo.html

It is ok to enter the test data while the test is running but the returned results should of course not be used as a basis to stop the test (I can explain this in more detail...)!

For people that like mathematics: the web page is based on the following script

http://hardy.uhasselt.be/Toga/sprt_elo.py

(I also have a C++ version in case people are interested).

[hagtorp]

Can I see the version in c ++?

[Michel]

Here

http://hardy.uhasselt.be/Toga/sprt_elo_snapshot.tar.gz

[hagtorp]

Thanks.

[Michel]

Hi,

If you want to use the C++ code for something please download again. I have removed a few lines of dead code in brownian.cpp (they were already removed in the Python version). There is no effect on execution.

[Dann Corbit]

That Brent root finder has been around for a long time. I remember playing with it 25 years ago.