Elo Style Models - My Thoughts

Skype: @TennisRatings

There has been quite a lot of publicity lately about Elo Style Models predicting match results, and considering that designing one of these was how I first started betting on Tennis, I feel quite well placed to make comment.

For those who don't know much about Elo Style Models, you can check out https://en.wikipedia.org/wiki/Elo_rating_system which gives an overview of how it works.

In essence, it's a rather accurate guide of working out player levels - a decent Elo model for Tennis, for example, will certainly rank Novak Djokovic and Serena Williams as the best players in the world, as do the current world rankings, and also as would the rankings if they were decided by combined hold/break percentage.

However, as with all models, one major issue comes with working out which inputs to use.  A high profile Elo model recently stated that they used two separate Elo ratings - one based on all a player's matches and one based only on matches on the relevant surface, and the two ratings were then combined to make event-specific ratings.

This approach is flawed in several major ways - a player's ability on a clay court, for example, is not necessarily an indicator of their likely performance on a hard court, or grass.  Can you imagine using an overall rating for the likes of clay courters Fabio Fognini, Pablo Andujar or Paolo Lorenzi on a hard court?  Despite Lorenzi's freak run at this year's US Open, it would be extremely incorrect.

Another unavoidable issue is the usage of entire career matches, as opposed to more recent ones.  Certainly at the very least, extra weighting should be given to more recent matches.  Usage of entire career matches is likely to over-rate declining old players, who will be able to still have a solid Elo rating based on their career exploits, as opposed to the fact they are not nearly the player they once were.

The final problem that any modeller faces, whether they are using Elo or not, is that a model does not take into account other factors relevant to the match-up.  

A basic model cannot flag up if a player has problems against left-handers, if they are 7-0 down in head to head matches despite being ranked at a similar level to their opponent, if they are a flat-track bully that beats players worse than them with extreme regularity, but cannot beat those better very often (e.g. Richard Gasquet or Tomas Berdych).  A basic model will also be unable to take into account the effects of injury or fatigue which clearly have major impacts on a player's likelihood of winning a match.

All of the problems mentioned above mean that it is very, very difficult for Elo models, or any basic model which doesn't make adjustments for relevant external issues, to beat the betting markets.  I managed it, and got restricted in many places pretty fast, but that was quite a few years ago now, when the betting markets were less efficient.  I think it would be much tougher now.

I am sure that some, less betting-minded, individuals, are now thinking 'why is beating the betting markets important?'.  

It's pretty simple - pre-match/in-play markets on Pinnacle/Betfair are the most accurate method of establishing a player's winning chances in a match, because they are - by and large - extremely efficient.  Bookmakers and professional gamblers/syndicates are not in the habit of making gross errors very often, and if they do, they tend to be corrected pretty quickly.

Therefore, if a model is producing different numbers to the betting markets, this would be down to one of two reasons:-

1) The model is wrong.
2) The model is right, and the market is wrong.

Obviously if point 1 occurs, then having an incorrect model is worthless.    
However, if point 2 occurs, then it is possible to beat the betting markets, and the model is great.  However, many Elo models don't seem to be able to beat the betting markets, and certainly not by even a reasonable (say 3%+) margin.  

Considering this, it would be very reasonable to think that if a model cannot beat a betting market, then it should be consigned to the scrapheap.  I understand that someone will be a betting market assessment on one high profile Elo model soon, and it will be very interesting to see the results over a large sample.

This is also the case for the various models that have cropped up in Cricket of late.  If you can beat the betting market, then the model truly deserves respect and praise.  However too many times, live on TV, we see absurd win or run predictions which clearly bear little relevance to the match state or likely future outcome.  Several of these models even state that they shouldn't be used in the betting markets.

If they can't do that, how can they be accurate, and what really is the point?
A Selection of TennisRatings Products for 2017
Please visit the
TennisRatings Products links for a full overview of our fantastic Tennis Trading tools, and the TennisRatings Subscription Packages link to see our great value range of discounted subscription packages!

Please check out our testimonials page!

The TennisRatings Daily Trading Spreadsheets have never been more popular!  

To find out more on how these can dramatically improve your Tennis Trading, check out the YouTube Video we made.

The Challenger Daily Spreadsheets cover all ATP Challenger Events and include projected hold percentages (for traders) and model prices (for bettors and traders).

Subscriptions are available for 3 months:-

The Lead Loss/Recovery Data Spreadsheets have taken the Tennis Trading World by storm - discussed in detail in October 2015 at the Matchbook Traders Conference these incredible spreadsheets highlight lead loss & deficit recovery in individual sets, as well as how often a player loses/gains the first break of the second set based on whether they won or lost the first set!