Hi,
Was just wondering if it could be considered a flaw that tennis trader doesn't take into account surface or gender.
For example breaks are more valuable on grass compared to clay. A break by a male player is more significant than a break by a female player.
As such Tennis Trader does seem to overestimate moves after a break of serve for a womens game on clay.
Any one else finding this?
Tennis Trader
Each sport you look to model has an underlying characteristic inherent within it. When you look at football it's goals and whom those goals are going to be scored by. So before the match you think hard about those two key questions. From that you can predict a number of key elements. All odds compilation works in the same manner.
On Tennis, Tennis trader looks at the odds and works out from the score and the odds what chance each player has of winning or returning a serve. From that seed it works forward to predict the odds. So the surface and gender isn't a factor it needs. If it looked at Murray vs Federer on grass the odds, and therefore seed stats, would differ from if the same match was played on Clay. The characteristics of grass and clay are different but that is contained in the seed value which is generated from the odds themselves. If you feel TT is wrong then you have to imply that the odds were wrong at the start of the match.
One thing that is impossible to predict is how these characteristics change during the match, but you can put some logic into that. In the Safarova vs Goerges match today Georges lost the first set unexpectedly. But if you look at the stats from that first set her first serve % was dreadful, there was a big deviation between the seed odds and what actually happened. At the start of the second set the market had discounted this below par situation so the market at the start of the second set would have varied from the original seed but you now had the chance to re-calibrate. However the question you now face is would Goerges get her first serve back on track, in which case you stick with the initial seed odds or do you think this is a characteristic of the match? In which case you reset to the situation at the start of the second set.
On this occasion I bet on Goerges to improve her serve in the second set and backed at good odds. She did and won the second set and her odds came crashing back in.
Hope that explains things and puts them in context.
On Tennis, Tennis trader looks at the odds and works out from the score and the odds what chance each player has of winning or returning a serve. From that seed it works forward to predict the odds. So the surface and gender isn't a factor it needs. If it looked at Murray vs Federer on grass the odds, and therefore seed stats, would differ from if the same match was played on Clay. The characteristics of grass and clay are different but that is contained in the seed value which is generated from the odds themselves. If you feel TT is wrong then you have to imply that the odds were wrong at the start of the match.
One thing that is impossible to predict is how these characteristics change during the match, but you can put some logic into that. In the Safarova vs Goerges match today Georges lost the first set unexpectedly. But if you look at the stats from that first set her first serve % was dreadful, there was a big deviation between the seed odds and what actually happened. At the start of the second set the market had discounted this below par situation so the market at the start of the second set would have varied from the original seed but you now had the chance to re-calibrate. However the question you now face is would Goerges get her first serve back on track, in which case you stick with the initial seed odds or do you think this is a characteristic of the match? In which case you reset to the situation at the start of the second set.
On this occasion I bet on Goerges to improve her serve in the second set and backed at good odds. She did and won the second set and her odds came crashing back in.
Hope that explains things and puts them in context.
I think it's quite an interesting question posed by mpendle and a very informative reply by Euler.
Perhaps to call it a "shortcoming" of TT would be more accurate than to call it a "flaw", but at least on the surface it seems it would be an interesting enhancement to be considered for TT.
Euler, I think I know what you're getting at when you say "The characteristics of grass and clay are different but that is contained in the seed value which is generated from the odds themselves", in that a player's ability on various surfaces and conditions will affect the odds the market places on them to win the match, which in turn, would drive the subsequent odds in the match on TT...
But let's say there are various matches where the odds are all at exactly identical odds to start to match. One is a men's match on clay, another is a women's match on grass, another is a men's match on hardcourt, etc.... going into the first game of the match, would TT predict a service break to cause the *exact* same change in odds for every one of these matches regardless of gender or surface?
If so, while it would of course be possible to "manually" compensate for it mentally, I think it would be very valuable if TT were to produce differentiated results automatically. I realize that it's impossible to factor in every single differentiating factor and create a filter for it - and at some point you have to draw a line - but I think those two factors are quite broad enough to be considered, and without doubt cause a very statistically meaningful differentiation in the stats.
Perhaps to call it a "shortcoming" of TT would be more accurate than to call it a "flaw", but at least on the surface it seems it would be an interesting enhancement to be considered for TT.
Euler, I think I know what you're getting at when you say "The characteristics of grass and clay are different but that is contained in the seed value which is generated from the odds themselves", in that a player's ability on various surfaces and conditions will affect the odds the market places on them to win the match, which in turn, would drive the subsequent odds in the match on TT...
But let's say there are various matches where the odds are all at exactly identical odds to start to match. One is a men's match on clay, another is a women's match on grass, another is a men's match on hardcourt, etc.... going into the first game of the match, would TT predict a service break to cause the *exact* same change in odds for every one of these matches regardless of gender or surface?
If so, while it would of course be possible to "manually" compensate for it mentally, I think it would be very valuable if TT were to produce differentiated results automatically. I realize that it's impossible to factor in every single differentiating factor and create a filter for it - and at some point you have to draw a line - but I think those two factors are quite broad enough to be considered, and without doubt cause a very statistically meaningful differentiation in the stats.
-
Nero Tulip
- Posts: 708
- Joined: Wed Apr 15, 2009 5:29 pm
Euler, I think your understanding of it is reasonable, but as a seasoned tennis model maker, I can tell you that I believe, though it's heart is in the right place, there are many instances when TT is wildly inaccurate. Perhaps I should attach to this a disclaimer that this is simply my opinion, and that I acknowledge that this is a difficult place for me to post this opinion. If the mods wish to remove this entry, I won't be offended, though I would hope I won't get a ban or warning.
To take the above mentioned example of differing surface; This is not (it seems) contained within TT, and is a massive factor. To state that you can derive the right odds simply from SP or worse still - current price, is to not fully understand how the tennis model works in terms of the critical input that governs such models. I won't flesh this out, as it seems BetAngel themselves have not posted much info on it, and so I'll leave it to them if they want to. Suffice to say, this 'critical input' is highly sensitive to surface.
An above poster mentions that you might have two identical starting prices on different surfaces. TT would populate it's forecasts at different scores to show the same odds. The reality is, the odds should be markedly different, especially when a player is ahead in a set.
There is much BetAngel might do to improve the model. I can tell you however that there is layer upon layer of detail you can add to this and still not get that close. The important thing is to use it as a very rough guide, learn from it's suggestions and form your own ideas about momentum / resistance around it. I don't need to say that anyway, as you always seem to be a very knowledgeable person on the forums.
To take the above mentioned example of differing surface; This is not (it seems) contained within TT, and is a massive factor. To state that you can derive the right odds simply from SP or worse still - current price, is to not fully understand how the tennis model works in terms of the critical input that governs such models. I won't flesh this out, as it seems BetAngel themselves have not posted much info on it, and so I'll leave it to them if they want to. Suffice to say, this 'critical input' is highly sensitive to surface.
An above poster mentions that you might have two identical starting prices on different surfaces. TT would populate it's forecasts at different scores to show the same odds. The reality is, the odds should be markedly different, especially when a player is ahead in a set.
There is much BetAngel might do to improve the model. I can tell you however that there is layer upon layer of detail you can add to this and still not get that close. The important thing is to use it as a very rough guide, learn from it's suggestions and form your own ideas about momentum / resistance around it. I don't need to say that anyway, as you always seem to be a very knowledgeable person on the forums.
There are an almost unlimited number of ways to model anything and an infinite number of variables. The trick with the software is to present it in a manner where you need a small number of variables to give you a decent model. It is possible to add many more layers but that becomes much harder to put into a sensible format for people to use.
If you have Federer and Nadal on grass and clay the seed odds will be different and that will be reflected in it's predictions. It is adjusting for different surfaces. Having tested this on grass, clay and other surfaces this has proved consistent. If you put in some odds it will work back to decide the supremacy between players and this is a reflect on the surface as well. If not then you can hypothesis that the Betfair prices are wrong.
From the moment the match starts though the dynamics will change and that is where the user needs judgement. It would be amazing if any model was able to accurately predict these dynamics without any further input or judgement. I think it's fair to say that clay matches will play out different to grass but the supremacy between players isn't that variable as that is discounted into the price.
If you have Federer and Nadal on grass and clay the seed odds will be different and that will be reflected in it's predictions. It is adjusting for different surfaces. Having tested this on grass, clay and other surfaces this has proved consistent. If you put in some odds it will work back to decide the supremacy between players and this is a reflect on the surface as well. If not then you can hypothesis that the Betfair prices are wrong.
From the moment the match starts though the dynamics will change and that is where the user needs judgement. It would be amazing if any model was able to accurately predict these dynamics without any further input or judgement. I think it's fair to say that clay matches will play out different to grass but the supremacy between players isn't that variable as that is discounted into the price.
-
Nero Tulip
- Posts: 708
- Joined: Wed Apr 15, 2009 5:29 pm
I do agree that no model is perfect, and it is impossible pretty much to be spot on all the time, the best one can hope for is to be close-ish on average. To find the middle ground in what is a very volatile environment.
Of course yes the odds will be different for same set of players on different surfaces, but TT does not reflect the necessary change to that input I was talking about. It does take into account the SP or current price, but that is not enough. There are two important inputs, one is the price, the other is the thing I am talking about.
TT does do something clever by adjusting using the current price, but you really have to take a big chance using that adjustment because some prices result in a crazy alteration in the underlying inputs. ie. The market can be a long way out, due to momentum, supply and demand etc.
Anyway, I do think it's a great tool, and you obviously get much from it and that's all that counts. My 2p on it for those new to it is to go easy on mid match adjustments, and form your own opinion on momentum both in the match and the market, getting both on your side at the same time results in the best trades.
Of course yes the odds will be different for same set of players on different surfaces, but TT does not reflect the necessary change to that input I was talking about. It does take into account the SP or current price, but that is not enough. There are two important inputs, one is the price, the other is the thing I am talking about.
TT does do something clever by adjusting using the current price, but you really have to take a big chance using that adjustment because some prices result in a crazy alteration in the underlying inputs. ie. The market can be a long way out, due to momentum, supply and demand etc.
Anyway, I do think it's a great tool, and you obviously get much from it and that's all that counts. My 2p on it for those new to it is to go easy on mid match adjustments, and form your own opinion on momentum both in the match and the market, getting both on your side at the same time results in the best trades.
Hello,
one thing on the predictions of TT I really don´t understand happened today in the match between Petzschner and Berdych. At some point the odds for Berdych were at 1.45 an he had to serve. The predictions from TT were 1.43 if B.would win the next point. But if he would win the game the predictions said 1.46 ??? This was the first time TT made predictions that I would call at least strange. Perhaps someone can give an explanation for this behavour or has seen something similar. Normally TT is a very usefull tool for me, but this time I had doubts if I should really trust the predictions for the rest of the game.
Greetings
one thing on the predictions of TT I really don´t understand happened today in the match between Petzschner and Berdych. At some point the odds for Berdych were at 1.45 an he had to serve. The predictions from TT were 1.43 if B.would win the next point. But if he would win the game the predictions said 1.46 ??? This was the first time TT made predictions that I would call at least strange. Perhaps someone can give an explanation for this behavour or has seen something similar. Normally TT is a very usefull tool for me, but this time I had doubts if I should really trust the predictions for the rest of the game.
Greetings
The most common mistake people make when using TT is to set up the match type incorrect. Several emails into support this week have been resolved through people using the wrong settings. I.e. Queens is a best of three set, French open was a best of five.
You can ever get any generic model spot on. All we can do is get near. There are so many variables we have to strike the balance between usability and what it can put out. We feel we have the right balance at the moment without over complicating it. Yes there are more inputs you could use and if anybody has some that they feel we would benefit from then we can add that, if you give us the detail and rationale.
When we coded the model we looked very long and hard at how to derive the start values and the market does an excellent job of that as it appears to be very accurate over time. When you start Tennis trader it seeds the model from the odds. So it looks at the odds and derives the chance of a player winning a point from that. The market is the seed. We went through all 10,000 combinations to see how best to derive what the market was telling us. If it feels Federer has a better chance of winning a point on grass against Murray than on clay the market will discount that into the start price and TT derives that value from there. We have considered giving the user more control over this process, so we may modify that aspect in the future. Again, so many variables.
But, say Murray has a 75% chance of winning a point then the chance of winning a game to love is .75^4. TT works through all the possible combinations of points to get a game %, then it works through all the possible game % to get a set % and so on. Calibrate resets this calculation according to the current score so it can work out how the market has shifted since it started. This is an added complication, the way the match plays out.
We looked at ten years of surface history at the number of games and sets on grass, clay and hard court when trying to work out if this had a big impact. This is because once we know the game % we need to factor in how many games there could be in a match and what the standard deviation is, to get reasonably accurate set and match predictions. The difference between Grass and clay is only 0.02 sets per match and typically a couple of games difference. Most of this difference is due to some of the extraordinarily long games that have taken place at Wimbledon over that period. If you strip out exceptionals most surfaces end up with similar numbers. So we didn't consider that to be a critical factor and that played out in testing.
When using any model you have to use your judgement on what it is telling you. What we have tried to do in TT is to say to you. "Hey, if you do this at this point you are running a big risk" or vice versa. This should allow you to make decent trading decisions through a match. It completely changed the way we looked at Tennis.
You can ever get any generic model spot on. All we can do is get near. There are so many variables we have to strike the balance between usability and what it can put out. We feel we have the right balance at the moment without over complicating it. Yes there are more inputs you could use and if anybody has some that they feel we would benefit from then we can add that, if you give us the detail and rationale.
When we coded the model we looked very long and hard at how to derive the start values and the market does an excellent job of that as it appears to be very accurate over time. When you start Tennis trader it seeds the model from the odds. So it looks at the odds and derives the chance of a player winning a point from that. The market is the seed. We went through all 10,000 combinations to see how best to derive what the market was telling us. If it feels Federer has a better chance of winning a point on grass against Murray than on clay the market will discount that into the start price and TT derives that value from there. We have considered giving the user more control over this process, so we may modify that aspect in the future. Again, so many variables.
But, say Murray has a 75% chance of winning a point then the chance of winning a game to love is .75^4. TT works through all the possible combinations of points to get a game %, then it works through all the possible game % to get a set % and so on. Calibrate resets this calculation according to the current score so it can work out how the market has shifted since it started. This is an added complication, the way the match plays out.
We looked at ten years of surface history at the number of games and sets on grass, clay and hard court when trying to work out if this had a big impact. This is because once we know the game % we need to factor in how many games there could be in a match and what the standard deviation is, to get reasonably accurate set and match predictions. The difference between Grass and clay is only 0.02 sets per match and typically a couple of games difference. Most of this difference is due to some of the extraordinarily long games that have taken place at Wimbledon over that period. If you strip out exceptionals most surfaces end up with similar numbers. So we didn't consider that to be a critical factor and that played out in testing.
When using any model you have to use your judgement on what it is telling you. What we have tried to do in TT is to say to you. "Hey, if you do this at this point you are running a big risk" or vice versa. This should allow you to make decent trading decisions through a match. It completely changed the way we looked at Tennis.