Lichess to FIDE rating converter

TLDR here's the converter: ethanlebowitz.github.io/RatingConverter/index.html

Hey folks.

I've always been a little curious about what my FIDE rating would be if I had one. Three years ago user dudeski_robinson did a linear regression analysis on around 3000 accounts to create a model for FIDE rating as a function of lichess blitz and classical ratings. His work can be found here: lichess.org/forum/general-chess-discussion/how-to-estimate-your-fide-rating-conversion-formula-inside

Things have changed since then (rapid time control hadn't even been introduced yet) and I wanted to update the work he did.

What I did differently:

- Bigger data
The original analysis was done with 300,000 accounts that yielded about 3,000 data points in which the user listed a FIDE rating. This analysis was done with a little over 1 million accounts that yielded about 28,000 data points with a FIDE rating.

- More models
I generated a total of 15 linear regression models so you can get the most accurate results with any combination of variables (variables being bullet, blitz, rapid, and classical ratings). That's one for every combination of variables. This way you no longer need to use exactly a blitz and a classical rating. You could use all four if you have them, or just one if that's all you have.

- A website
So you don't have to pick the right model and apply it by hand.

Note: one thing that hasn't changed is that this is still a very very rough estimate. Take it with a hefty grain of salt.

If you want to read about the specifics about how this project was done you can here: ethanlebowitz.github.io/RatingConverter/about.html
Here's the source code for the website: github.com/EthanLebowitz/RatingConverter
I'll likely upload the code for some of the python/r scripts I used for data mining, parsing, and processing once I've cleaned them up a little.

Let me know what you think!

so you're 1337 :)

Haha I guess so

This is an interesting project, but as you yourself said, one must take it with a grain of salt... especially when you are trying to correlate ratings at very different time controls, where the correlation is VERY loose. No number of regression models can compensate for this.

If it is FIDE rating you are trying to model, the best method is to ONLY compare the closest possible time control -- i.e. Classical. The correlation is tighter in this case.

If you are using other time controls, it would be a good idea to also calculate the standard deviation or 50% confidence interval in the corresponding FIDE rating. In general, I predict that you will find that the shorter the time control, the wider the confidence interval will be.

You make some good points phoenixshade. This is definitely more of a novelty than anything else. There is a large degree of error involved.

I don't totally understand what you meant in your last paragraph, but the other time controls don't seem to have a much weaker correlation with the FIDE rating than the classical rating does, though none of the correlations are extremely strong. The Pearson correlation coefficient for FIDE as a function of classical is 0.605, and 0.538 for FIDE as a function of bullet. That's not all that much lower.

Are you doing some kind of polynomial regression? I'd recommend probably 3rd or 5th order. If the correlation is non-linear but you are using linear regression techniques, the Pearson coefficient will be high.

I'm really surprised that there is not a MUCH greater difference between the Pearson coefficent for bullet vs. classical. My first thought as to the explanation was the above.

Also, I don't know if you can do this, but maybe prune your dataset to exclude lichess RD>55, as well as FIDE ratings whose last updates are far removed in time from the lichess rating. Perhaps you could use their lichess rating AT THE TIME of their last FIDE rating change.

I think by selecting the best dataset, you can drastically tighten the classical lichess/FIDE correlation. Maybe even reduce the size of the grain of salt.

Magnus is 2470 lol

Haha, ya lots of error. Like I said it's a very very rough estimate.

@phoenixshade thanks for the advice. I'm definitely new to data analysis and this project was largely to help me learn. What do you mean by RD>55? I like the idea of using the lichess rating at the time of the last FIDE change, though I'm not sure if I can readily access the information necessary for that in bulk. I'm probably not going to do any more work on this project in the near future because I'm a little burnt out, but I'll definitely take these things into consideration if I revisit it, and might message you to discuss it if that's alright.

my classical rating is 1400 so what rating can reach FM?