Sportpunter's Tennis Model

The Rating Method

The first step in creating a model that can predict tennis matches as accurately as possible is to produce a rating system. Whilst many punters would use the official ATP tennis ranking’s, there is a lot of evidence that suggests that the ATP tennis ratings are not a good indication of current form.

Clarke (1994) produced what is called a SPARKS method (short for set-point-marks) which calculates the margin of victory of a tennis match.

A margin of victory is important when looking at a rating system. Bedford and Clarke (2000) found that the SPARKS method produced significantly better predictions than the official ATP tennis ratings. Their method though was relatively simple despite being amongst the first to look at this area. They didn’t look at surfaces and many other factors, however it was a definite step in the right direction.

In the Australian Open 2002, there was a lot of media coverage about why most of the top seeds were eliminated from the event early on. In a newspaper article in the Australian Financial Review, written by head of Champion Data and myself, we outlined that a lot of the top seeds were not in the best of form especially on the hard court surface.

So why isn’t the ATP ratings a good predictor for tennis matches?

There are a number of reasons why.

• The ATP ratings do not take into consideration the quality of the opposition.

If Lleyton Hewitt defeats Pete Sampras in the first round of the Australian Open, he would gain just as much as if he had defeated Jakub Herm-Zahlava.

• As mentioned previously, The ATP ratings do not take into consideration the margin of victory that the player won or lost by.

Lleyton Hewitt would gain just as many points in defeating Pete Sampras 6-0 6-0 6-0 as he would if he had defeated him 6-4 2-6 7-6 0-6 10-8.

• The ATP ratings give points for players who win due to the opposition retiring.

If Lleyton Hewitt was behind 2-6 2-4 and then Pete Sampras retires, Hewitt would receive points for progressing to the next round.

• The ATP ratings system gives players points for walkovers.

If Pete Sampras obtained an injury between matches and could not front up for the next game against Hewitt, Hewitt would receive points for progressing to the next round despite not playing a game.

• The ATP ratings system does not take the playing surface into account.

Many players play better or worse on certain surfaces and this has to be taken into consideration when looking at a players performance. Defeating Pete Sampras on grass is a lot better win than defeating him on clay.

• Probably most importantly, the ATP ratings system does not look at current form.

According to the ATP ratings system, the last 12 months of tennis is taken into consideration for their ratings. This means that a players performance 12 months ago has just as much weighting as what that player did last week. As a good example of this, Guestavo Kuerten was ranking number 2 at the end of 2001, however he had lost his last seven games.

• The ATP ratings system does not take into considering any home ground advantage
• The ATP ratings system does not take into consideration how long it has been since a player played on a particular surface
• The ATP ratings do not account for head to head performances.
• The ATP ratings system does not take into consideration other important tournaments like the Davis Cup, challenger and qualifying results.
• The ATP ratings system gives bonus points to players with good sportsmanship and other factors which is hardly representative of their current form.

Our Ratings System

Sportpunter's ratings system does take all this into consideration and awards players with a rating based on the SPARKS method. Each player also has a surface rating which is added to their overall rating to get an expected marks victory on each particular surface. These ratings are added together using linear regression. Based on each weeks performance, a players ratings might up or down which is shown here for men and here for women. Likewise their surface rankings might also change as shown here for men and here for women. The information on player preferred surfaces is only given by rank. The amount which a players rating changes based on the outcome of a match is found by a statistical method called ‘exponential smoothing’ which changes the ratings by a percentage after each match.

Interestingly a couple of factors were also considered. One is the head to head approach. Many punters look at past head to heads to predict what is going to happen in a current match. However punters will generally only look at how many wins each player has whereas they should be looking at their current form when they played head to head and their margin of victory. For a hypothetical example, Ferrera might have played Hewitt 5 times head to head, Ferrera might have won the first 4, but Hewitt the last game. Does Ferrera currently have the edge over Hewitt based on his 4-1 head to head record, or was it Ferrera's win came about when he was number one in the world and Hewitt was only just starting in tennis.

Sportpunter's tennis model takes these into consideration, by looking at the expected margin of victory and the actual result based on the quality of the players at the time, the surface it was played on and many other factors.

Home ground advantage in tennis, whilst despite smaller than many other sports still exists. Sportpunter has added this into their model. By a full analysis of every single player played in every tournament around the world, Sportpunter has been able to determine exactly how much home ground advantage comes into play and how much of this is just merely due to the court surface and other factors.

The theory that players will not come back well after playing a five set match however does have an effect. It has been shown statistically that players do tend to play below their usual performance if they last played a five set match in grand slam tournaments. Sportpunter's model accounts for this. And has accurately determined how much of a factor it has.

Tiredness and unfamiliarity are other factors which Sportpunter's tennis model takes into consideration. If a player is "first-up" on a particular surface, Sportpunter has determined what effect this will have on their performance on average. Likewise Sportpunter's model looks at how many matches a player has played on a surface throughout the tournament.

Sportpunter's predictions are updated every day, taking into account every match previously played.

One of the advantages of a computer model for analysing tennis is it's memory. While humans will forget how a tennis player played one month ago, a computer model will never forget. And more importantly it will determine to what degree that match a month ago has an effect.

Head to Head Probabilities

The expected outcome of each match can easily be converted to a probability. The method to convert these expected marks to probability is done by another statistical procedure called ‘logistic binary regression’.

Calculating the Tournament Probabilities

Given that we have the probabilities for head to head matches, we can simulate the entire tournament. The total amount of possibilities of a tournament are incredibly high. For a small 32 player tournament, there are a total of 31 matches, and therefore there are 2³² possible outcomes which is over four billion different outcomes. Given this, the best way to calculate the probabilities for a tournament are via computer simulation. Each match is simulated and their rankings are adjusted after each round. This is an important step because if a little known player with a small ranking previous to the tournament made the final, his ranking at that time would be a lot higher originally due to his good form throughout the tournament so far.

So the tournament is simulated approximately 10,000 times depending on the size and the number of matches remaining.

Using the model to gamble

Original this was not the purpose of the model, it was just for matter of public interest, but seeing if the model is profitable is an important part of any statistical model when predicting sport outcomes.

Why is this? Well it’s quite simple. An important distinction to make is that bookmakers do not make odds based on the probabilities of winning, but rather what the general public thinks the probabilities of winning are. Their main concern is to balance the books based on what the average joe-bloe believes.

Therefore the model could be proven a statistically better predictor than the general public if it is profitable based on bookmakers odds. Given below is a step by step method of how one can gain an advantage over bookmakers and have the potential to make money by gambling on tennis matches.

The Gambling Technique

Converting Bookmakers Odds to Probabilities

By converting a bookmaker's odds to probabilities we can directly compare these to our own probabilities to see if there is a possibility of an advantage in a gamble. The inverse of the bookmaker's price is the expected probability.

For example in late March 20001 the bookmaker's gave Fernando Gonzalez (CHI) odds as high as $4.50 to defeat Pete Sampras (USA). This means that the bookmaker's (or the general public) believe that Fernando Gonzalez (CHI) have approximately a chance of winning. We had predicted a 71.7% chance for Pete Sampras (USA) to win the match, consequently this means that Fernando Gonzalez (CHI) have a 28.3% chance. This probability is higher than what the bookmaker's have Fernando Gonzalez (CHI) at and therefore this is where we have an advantage over the bookies and would gamble on Fernando Gonzalez (CHI) for this game.

Put simply, we have a 28.3% chance of returning $4.50 from a $1 bet, so on average our $1 bet will return 0.283 * $4.50 = $1.27. Hence an expected profit of 27%.

How much of an advantage do we have?

The advantage over the bookmaker's, or overlay, is calculated by taking the bookmaker's price into account by the following formula:

Overlay = [Our probability * Bookies Price] – 1

Therefore in this game, we had an overlay of (0.283 * 4.50) - 1 = 27.3%

This overlay is very large, and represents a very good betting opportunity, even though we still believe for Pete Sampras (USA) will win the match.

It is important though not to bet on any match that has a small overlay. To take into account some error, one should only bet on matches where a large overlay is recorded.

Not all matches will we have an advantage over the bookmaker's however. If the bookmaker's price is similar to our probabilities then there is no room for an advantage. This is mainly due to the fact that the bookmaker takes a 5% to 8% overlay per game.

How much should be bet?

Even if the odds are on your side, you still need to guard against losing all your bank. We can work out mathematically the percentage of your bank you should bet to maximise your rate of growth.

The amount to bet is given using a system called the 'Kelly' method which was found by Kelly in 1956. It uses the bookmaker's price, your probability and the amount of overlay that you have in determining how much to gamble. For more information on money management please see our money management page.

When shouldn't we bet?

Although some will say this is up to the individual, I believe there are a few times when one shouldn't bet on an event. One of these is when a player has not played many games on a particular surface. When the matches are shown from our website, there is also a column that shows 'Games (p1)' and 'Games (p2)' which refer to how many games player one and two have played on the current surface that the tournament is being played on. Betting on matches where these values are low could result in more long term losses. The reason for this is is that when a player starts his first game, he is given a surface rating of zero, this is because we have little information about this player. However he might be a clay court specialist, but our ratings will not reflect this. Therefore it is only recommended to gamble on a player once he has played several games on that surface and can develop a satisfactory surface rating.

Another time in which one probably should not gamble is when a player is returning from injury or currently has an injury. Likewise I would refrain from gambling is a player retired recently from singles or doubles matches.

There are also times when in the past it has been proven statistically that betting on certain matches might not be profitable in the long term. This is due to a number of factors including a bias towards the underdogs with bookmakers prices. To see a full analysis of the tennis model then click here.

When does Sportpunter bet?

Sportpunter, follows the bets and bets on exactly what the model suggests. This way you can feel secure in that we want to make the model as good as possible because we too are betting on it. Our suggested bets that we display are for when both players have played at least 5 games on that particular surface. We don't recommend betting on games where one of the players has played less than 5 games on the surface.

Similarly we look at overlays when making our bets. History has shown that there is little value in betting on big outsiders in tennis; the favourite win in tennis more often when compared to other sports. Bookmakers know this and because of the favourite long-shot bias we don't bet on players where the calculated probability of then winning is less than 30%. When a player has a probability of winning between 30% and 50% we only bet on them if the overlay is above 25%. But when we calculate the player as a favourite, we bet on them no matter what the overlay is.

This is what comprises Sportpunter's tennis selections, and is exactly what goes into out betting history. The Excel spreadsheet that you have received has a system where you can judge the minimum overlay to bet based on the probability. It is currently more "smooth" than the simple method mentioned above and might be more to your tasting. If not then simply change it so that your bets will equal the one's that Sportpunter suggest.

Where did you find those odds?

In fact, well we didn't. Because different tennis matches are played each week bookmakers can only suggest odds the day before they are being played. This means that bookmakers often release odds close to the start of the tennis match. Whilst we would like to use these odds to help you with your suggested bets, many people like to know far in advance who they should be betting on. Hence we create "average odds" that are standardised over the average of all bookmakers to 103%. Pinnacle Sports has about that same margin, and betfair have as low as 102% and even less if you know how to lay. Hence the odds given for the suggested bets are more like average odds. You should be able to find odds that are better than the one's suggested.

How have we gone so far?

Our past record for ATP tennis is shown here. From this it can be seen that in only half of the year 2003, we managed to increase our bank account from $500 to over $10,000. Our %ROI or %return on investment has constantly hovered around the 5% mark.

Given our good long term history, the potential to make money is very high. There are 3,000 ATP matches in a year and 2,500 WTA matches. Hence lets suggest that we bet on 1000 matches. The average betsize has been shown to be 20% of your bank. And with the quarter kelly method of betting this means an average bet size of 5% per bet of your bank. Lets assume that you have a bank size of $5,000. Given this one can approximate how much they believe they will earn in a year.

Potential Profit = Bank Size * %BankBet * #Games * 10%ROI

Potential Profit = $5,000 * 5% * 1000 * 5% = $12,500.

So with a modest bank size of $5,000, one could gain a profit of $12,500 should the results in the future follow that of the past. And there is no reason why it shouldn't. Although the betting history has been updated since April 2003. Long term followers will know that the model has been running since Jan 2001 with equally as good results.

And Finally…

I hope that you enjoy my website and get the most out of it for yourself. Whether you’re a punter, or just interested in tennis, or maybe interested in sports statistics and mathematics, I’m sure that you will get something interesting out of this website.

If you have any questions, please feel free to email me at jlowe@sportpunter.com. Otherwise happy punting!