Few queries

**antonyp22** · 10-14-14, 01:29 AM

When you talk about "percentage" are you referring to your pure win percentage?

**Baeoz** · 10-14-14, 02:42 AM

<br>
<br>
yeah. I mean the percentage of wins on all matches played during a given week. I figure if I predict this accurately it'll have a strong correlation with bets I make. My prediction rate drops off by a couple of percent for men from lowest to higher levels. It doesn't drop off so much for woman and there might still be an edge there on WTA.

It seems in ATP the favourite of the bookies doesn't lose as frequently, combined with my drop off in prediction means no positive expectation over all ATP.

**antonyp22** · 10-14-14, 03:19 AM

What you're experiencing is probably just that the more liquid markets (ATP and WTA) are far more efficient than the lower levels of tennis and that in order to maintain the same edge you'll need to introduce some more sophisticated modelling techniques/get better data

**Baeoz** · 10-14-14, 03:48 AM

Originally posted by antonyp22

What you're experiencing is probably just that the more liquid markets (ATP and WTA) are far more efficient than the lower levels of tennis and that in order to maintain the same edge you'll need to introduce some more sophisticated modelling techniques/get better data

That makes perfect sense. That's why I'm looking at regression/monte carlo/graph theory/kitchen sink.

I just wonder why my actual pure prediction rate drops off a few percent, on average, on these markets. I assume it's because they don't have as much 'mixing' between the main tours and the low levels.

**Baeoz** · 10-14-14, 04:04 AM

What's a ballpark prediction figure that I should aim for?
I only bet on underdogs that my model says will win, and because they're over $2, I make by on ITF as I get over 50% of these bets correct.
I'm investigating if I should invert that on high level, but I don't want to bet on $1.50 or less as you need 2 wins to cover a loss (>%0.66) instead of 1.

**antonyp22** · 10-14-14, 05:33 PM

Your pure prediction rate is an interesting one, perhaps you a) don't have enough ATP/WTA data b) haven't allowed for variables that are significant at an ATP/WTA level but aren't significant at lower level tennis

I think rather than limiting what odds you do/don't bet on, you're better off coming up with a way of figuring out what you believe is the true probability of a player winning which you can then use to price a H2H market. This way you may still find that betting on $1.50 odds can be +EV in certain scenarios.

The best way to do this would be to perform logistic regression or as you said perform Monte Carlo simulations on individual match-ups.

**Baeoz** · 10-14-14, 08:40 PM

Originally posted by antonyp22

Your pure prediction rate is an interesting one, perhaps you a) don't have enough ATP/WTA data b) haven't allowed for variables that are significant at an ATP/WTA level but aren't significant at lower level tennis

I think rather than limiting what odds you do/don't bet on, you're better off coming up with a way of figuring out what you believe is the true probability of a player winning which you can then use to price a H2H market. This way you may still find that betting on $1.50 odds can be +EV in certain scenarios.

The best way to do this would be to perform logistic regression or as you said perform Monte Carlo simulations on individual match-ups.

Cool. That makes sense. I've been conservative, backing only underdogs because I know that there's +EV on those and I don't need to guess. Now, to do a monte carlo on the data I have, what is a quick outline. As I understand Monte Carlo, you use a random number generator and if the number is below a probability you call that a win (for example) and above a loss. I can do that with the points in a tennis game, and when I do 10,000 iterations with probabilites for two players I unsurprisingly get the same result as when I use an analytic solution with the same probabilities. I'm not sure how I'd use it with betting data. I mean, do I use known odds (which I have) for the results of known matches then randomly run against those matches crediting wins and debiting losses based on the random number? I mean say I have a match with favourite $1.50 and estimated prob of 70% against underdog of $3 with prob of 30%. For each iteration the random number pops up and if it's less than 70% (0.7) I credit the $1.50 and if it's between 0.7 and 1 debit a unit? And so on for all the selected odds data?

If that makes no sense, it's probably because I'm a bit confused at how to apply monte carlo and regression in some situations.

Regarding ATP/WTA data I don't think I'm missing any. I'm quite keen on making sure the data is up to date and accurate. My win probability drops right off when it isn't as I weight newer games more heavily than older.

Thanks for your help too.

**antonyp22** · 10-14-14, 10:34 PM

The way I have used Monte Carlo simulation in the past is when I have had multiple data points that fit some sort of distribution. The distribution is then used as the input into the Monte Carlo simulation, I'm not sure if you're familiar with the Brownlow Medal, but that's how I carry out my analysis on that. I use a specific distribution for each player's predicted votes and carry out a simulation of the season based on each individual distribution.

Here, I think you're better suited to using a logistic regression if we're speaking in terms of purely predicting the probability of a player winning a tennis match, simply because you are producing ONE number for win probability and not 20 win probabilities so as to create a distribution.

I use an Excel add-in called "@RISK" for all simulations/distributions, PM me if you want to continue the discussion.

**Baeoz** · 10-14-14, 11:03 PM

Originally posted by antonyp22

The way I have used Monte Carlo simulation in the past is when I have had multiple data points that fit some sort of distribution. The distribution is then used as the input into the Monte Carlo simulation, I'm not sure if you're familiar with the Brownlow Medal, but that's how I carry out my analysis on that. I use a specific distribution for each player's predicted votes and carry out a simulation of the season based on each individual distribution.

Here, I think you're better suited to using a logistic regression if we're speaking in terms of purely predicting the probability of a player winning a tennis match, simply because you are producing ONE number for win probability and not 20 win probabilities so as to create a distribution.

I use an Excel add-in called "@RISK" for all simulations/distributions, PM me if you want to continue the discussion.

I haven't been posting enough to PM you yet.

Here's something like what I was going to PM:

I just run my model against data since last April and it 'bets' on those that match the criteria (predicted winner with +$2) and either scores the profit or notes a loss. This is grouped by dollar amount and also tournament level. EV and Z-score are calculated. I then just look at which tours/levels have the better EV.

I would like to do something with regression or monte-carlo. Just not sure what it is I'm supposed to do change and what stays the same. It's not really a tournament where I'm plugging in values at the start, as my model is updated each week. Though I could run each week again, but the winners are the winners, so I can't change that. I can only vary my presumed edge and the odds within the recorded window I think.