Saturday, May 18, 2013

How accurate is the USTA's Dynamic NTRP at predicting tennis matches?

There is an interesting debate on the Tennis Warehouse Talk Tennis forum about whether or not the NTRP system can predict matches.  While I weighed in early to say the overall system wasn't designed to do predictions, I said this because the goal of the system is to promote competitive play and the detailed ratings aren't published to facilitate predictions.  The truth is that the algorithm itself in essence does predict a match result and makes adjustments to each player's rating based on how much variance there is from the prediction.  Note that the system uses game differential, not win or lose so it doesn't actually predict who will win or lose a match but who will win more games.

Now, doing this prediction does require having the actual dynamic rating, and not just the level the players are playing at, so my Estimated Dynamic NTRP is an excellent way to see how well the "predictions" fare.  For our sample, I'm going to use some matches played at the Northern California Sectionals going on this weekend.  I've done some analysis for one of the teams playing there so have already been looking at the matches in some detail.

In one team match there were the following results:

  • A 4.77 beat a 4.40 6-2, 6-0 - Pretty much on target and expected
  • A 4.51 beat a 4.45 6-0, 4-6, 1-0 - Ratings predict a close win and it happened
  • A 4.57/4.37 beat a 4.67/4.37 7-6, 6-4 - A close match predicted, but went to the lower rated team
  • A 4.47/4.52 beat a 4.30/4.04 6-1, 6-1 - Pretty much expected
  • A 4.52/4.64 beat a 4.36/4.40 6-3, 7-6 - Also pretty much expected.
So my estimated ratings went 4-1 and the miss wasn't radically off.

Another team match had the following results:
  • A 4.81 beat a 4.09 6-3, 6-3 - Expected a worse beating, but still got it right
  • A 4.66 beat a 4.55 7-6, 2-6, 1-0 - Got the win right, not quite like would have been predicted though
  • A 4.40/4.69 beat a 5.05/4.15 6-3, 1-6, 1-0 - The lower rated team won so a miss but the higher rated team did win more games so the algorithm was right in a way.
  • A 4.61/4.43 beat a 4.49/3.99 6-3, 6-7, 1-0 - As predicted
  • A 4.64/4.61 beat a 4.51/4.52 3-6, 6-2, 1-0 - As predicted
Again, my ratings went 4-1, the miss actually being "right" on game differential though.  Also, matches with that big a gap are a wildcard as if the more even team can find the weaker player, the advantage is nullified.

Overall, 8-2 in predicting matches.  I'll take that.

Look for more on the other matches later.