Multiple matches per player per day with Kalman

[scotty779]

Does the Kalman model not work well when you have players with multiple matches at the same time period? I'm struggling to get any sensible ratings/predictions from the Kalman model after trying many different parameter combos. I had no trouble with Glicko on the same dataset.

[SamDuffield]

Hi! Great to see your interest in SSM rating systems!

I did some quick sims which I've pushed to a new branch here https://github.com/SamDuffield/abile/blob/glicko-kalman-same-time/simulations/glicko_kalman_same_time.py

Here I compare Glicko and ExKF on 1000 simulated match results over 10 players (~200 matches per player) all at the same time. Importantly I used the static parameters from the WTA tennis training.

The results look remarkably similar for Glicko and ExKF:

Final posterior skills:

Match predictions based on final posterior skills: