A dog system but is it worthwhile ? - #1 Horse Racing Forum in Australia for Race-Horse News

ATipster

Open
user 1491
Offline

Offline
Alias: None
Posts: 106

Original Post 2011-May-18, 01:24 PM

All,

I have created a program to download data from thedogs.com.au and analyse a dogs form to compute a time at todays distance and track.

The program essentially analyses the time to determine the dogs likely average time at this distance/Track and then determines a standard deviation for a likely minimum and a likely maximum time the dog could run. It then compares each dog to the other dogs using a simple monte carlo simulation running the race 10000 times and works out the probability of the dog winning. Of course this probability can then be set to a price.

Using this complex rating program and selecting the dog with the best time (I have analysed 11 months of unitab data) returns 3240 selections for 795 winners (24.54%) and a level stakes loss of 3.8% (121 units). Over the same period unitab 100 raters returned a loss of 21.5%. So I am guessing my rating method is not too bad.

By applying 2 filters I can weed out over bet selections (box 1 and last start winners) so that only 908 selections are used. These return 19.32% profit at a strike rate of 18.83%.

These 2 filters are pretty standard overbet situations as even adding the to the unitab top rater improves the loss from -21.5% to a -13% loss.

If I apply a value filter (price higher then estimated chance) I can further reduce this to 305 selections for 75 winners (25%) and a 37% profit.

Which of these methods would you recommend to bet ? The first with 908 selections but it can have dogs running around at $1.10 through to $150. Or the value filtered selections ?

On $1 bets the 908 selections return $175.50 where as the value filtered selections return only $113. I am mindful of the maximum I would likely to be able to get on (in the future) without affecting the odds and this is leading me to the 908 selection method. Also it had 2 losing months whereas the value selections had 3 losing months. This seems obvious to me as the value filters selections will be of a higher price and a lower strike rate.

Any thoughts are appreciated.


	Logged

fours

Group3
user 704
Offline

Offline
Posts: 946

2011-May-18, 01:37 PM

A Tipster,

Back all positive expectation scenarios.

The % of bank you bet should vary with strike rate expectation and or you can use full or partial Kelly. In part the correct answer depends on your own comfort zones as most people cannot stomache full Kelly.

Fours


	Logged

Wenona

VIP Club
Group 2
user 175
Offline

Offline
Posts: 4352

2011-May-18, 01:52 PM

From what you write, it would appear your non-value selections (ie those you think are unders) are returning about a 10% flat stake profit.

You obviously have a formula to convert your ratings to required odds.

The first thing I would do would be to work on that formula so it's estimated required odds are closer to the actual winning percentages (ignoring $ returns)

It would seem you may be systematically underestimating the winning probabilties of your top selections.

Once I was happy that I was converting ratings to required as accurately as I could, I would only bet on 'value' selections. This approach would also mean that you don't have to eliminate all the Box 1's or last start winners.

Out of interest, do you show a profit on the Box 1 and last start winners who you currently have as value?


	Logged

Wenona

VIP Club
Group 2
user 175
Offline

Offline
Posts: 4352

2011-May-18, 01:57 PM

I also agree with fours about having some methodology around your investments that took into account risk and edge.

Some type of modified Kelly would be good. I'd recommend you set some parameters though, if you have something priced at $1.50 and the market has it $12.00 I'd suggest you modify your bet to some predetermined percentage limit and accept that your assessment may be flawed to some extent.


	Logged

ATipster

Open
user 1491
Offline

Offline
Alias: None
Posts: 106

2011-May-18, 02:08 PM

Quote

Out of interest, do you show a profit on the Box 1 and last start winners who you currently have as value?

Box 1 shows a large loss of 11% but on only 40 selections.
Last start winners show a small loss of 4.8% on 138 selections.

So most of the selections that are being excluded by those filters are just under valued. I would think the small number of selections in box 1 / last start winners is the reason they may be showing a loss rather then a profit.

I was thinking of kelly staking but hadn't got that far into how the kelly stake would change the profits. Also I would need to understand the kelly stakings impact on the odds if I was to bet with the tab.


	Logged

ATipster

Open
user 1491
Offline

Offline
Alias: None
Posts: 106

2011-May-18, 02:09 PM

BTW fours and Wenona thanks for the replies.


	Logged

Wenona

VIP Club
Group 2
user 175
Offline

Offline
Posts: 4352

2011-May-18, 02:33 PM

I'd analyse the results across bands eg Top pick - second pick - third pick etc and also percentage bands 0-5%, 5-10% etc perhaps 8 bands that have about the same number of selections and see if this sheds any light.

By analyse I mean compare the total winners in each band against the total of the runners estimated percentages in those bands. This should give you a good indication if you can improve your price assessments from tweaking your formula.


	Logged

ATipster

Open
user 1491
Offline

Offline
Alias: None
Posts: 106

2011-May-19, 09:16 AM

Thanks for the top advice. I downloaded a few more months of data last night and also fiddled with my rating method. I was definitely under estimating certain price ranges and a few minor changes have changed the value method to show a 268.80 unit profit over 952 bets (28%). Where as the old method was showing a 141.1 unit profit over 417 bets. Althought the second method still has a higher profit on turnover if you look at the expected vs actual winners you get the following:

New rating method + Value = 151 winners vs expected 114
Old rating method + Value = 102 winners vs expected 75

To do the adjustments to the ratings I used the first 6 months of results I had and then confirmed it over the rest of the data I had.

Analysis of the results show a Chi test results of

New rating method + Value = 0.000221132
Old rating method + Value = 0.000576125

I might start monitoring both of these ratings methods real time for the next few months.

I would post the selections here but the unitab site doesn't put up greyhound races the day before so I can only post after the event selections for some of the races as I would not be running the results until each night and some races will have already been run.


	Logged

Wenona

VIP Club
Group 2
user 175
Offline

Offline
Posts: 4352

2011-May-19, 10:47 AM

So your actual winners are still 32% above your expected strike rate?

How do you calculate your expected prices?

Can you put up a couple of race examples so we can see the math?


« Last Edit: 2011-May-19, 10:49 AM by Wenona »	Logged

ATipster

Open
user 1491
Offline

Offline
Alias: None
Posts: 106

2011-May-19, 12:58 PM

Quote

How do you calculate your expected prices?

Its done with a Monte Carlo simulation. The only real inputs to it are the avg time, and the standard deviation of the dogs times. I can fiddle with the inputs to these based on box position, last run time, etc. At the moment I am only taking into account the last 2 recent runs and the box position changes. I am not taking into consideration track statistics which I think may be useful but am yet to get that data. I also think split times will be useful but have not implemented them yet.

To determine the chance of a dog winning it runs the race 10000 times. It uses a few random variables which are skewed slightly to increase interference, box position, etc. It determines a new dog time for each race based on the avg time and std deviation of a dogs time. If the dog has had only 1 run it will assume a deviation based on the average of all dogs at that distance which is usually in the range of 0.1 to 0.4 secs.

As far as examples here is one from yesterday (of course I picked an example of where I got the winner at value

):

Richmond Race 5 - 535m
Dogs were rated in the following order with the expected prices:

BOLT TED BOLT	$2.82
ANNABELLE	$3.46
ZIPPING JEWEL	$4.92
DIRECT DOLL	$22.28
QUEEN OF RAIN	$22.32
THOMAS VON	$46.11
BALUGA BOY	$47.01
TAUNT	$48.88

Total market should be close to 100% although rounding may have it slightly above or below.

The winner of the race was BOLT TED BOLT at $7.90 on unitab. (picture of finish is available at http://www.thedogs.com.au/Racing/RaceFinishImageDisplay.aspx?raceId=1002863)

I determine expected winners simply by dividing 1 by the price. So for instance a selection starting at $2 is 0.50 expected winners. I know this doesn't take into account takeout but I think this is a more conservative way of working it out. With a 15% take out the expected winners of a selection starting at $2 would actually be 0.425

If my math is wrong please let me know.


	Logged

arakaan

Group 1
user 192

Offline
Alias: was a TV show starring Jennifer Garner
Posts: 7879

2011-May-19, 04:45 PM

ATipster writes ... 2011-May-19, 12:58 PM

Richmond Race 5 - 535m
Dogs were rated in the following order with the expected prices:

BOLT TED BOLT	$2.82
ANNABELLE	$3.46
ZIPPING JEWEL	$4.92
DIRECT DOLL	$22.28
QUEEN OF RAIN	$22.32
THOMAS VON	$46.11
BALUGA BOY	$47.01
TAUNT	$48.88

I'm more intrigued with how your system came up with the 1 dog as the fav. on ratings in that race. chin


	Logged

ATipster

Open
user 1491
Offline

Offline
Alias: None
Posts: 106

2011-May-19, 06:27 PM

Lets just look at the top 3 rated dogs.

The times before any adjustments:

Raw best time at distance:

BOLT TED BOLT 31.04
ANNABELLE 30.68
ZIPPING JEWEL 30.59

Standard deviation for each dog:

BOLT TED BOLT 0.35
ANNABELLE 0.405
ZIPPING JEWEL 0.479

Add in adjustments for the box they are running from and you will note the difference between box 1 and box 4 winning times is approx 0.3 anyway. Just taking these things into account above you get:

BOLT TED BOLT could run as fast a time of 30.7 or as slow as 31.4
ANNABELLE could run a time as slow as 31.1
ZIPPING JEWEL could run as slow a time as 31.1

Of course Anabelle and zipping jewel could have run a faster time. But after you factor in interference,box position timings and recent performance/improvement you get a final time range for all dogs. This could have resulted in a larger or smaller time range for each dog.

The final difference between BOLT TED BOLT and ANNABELLE was only 6%. That's only 6 races more in every hundred which is not much of a difference.

Of course at the time I only knew it was the dog selected as the top rater and a price. I had to rummage through the times/figures to work the above out manually so there may be some issues with my manual calcs but it should be close to how I worked out the times.


	Logged

Gintara

VIP Club
Group 1
user 16

Offline
Posts: 5870

2011-May-19, 08:01 PM

ATipster writes ... 2011-May-19, 12:58 PM

Although different for each track, on average from memory leaders win 75% of races, so I'd say that data should feature fairly highly in your criteria if you can get it.


	Logged

ATipster

Open
user 1491
Offline

Offline
Alias: None
Posts: 106

2011-May-20, 09:20 PM

Well I started officially starting to follow my selections with small bets today.
I won't post the selections unless there is a noteworthy winner. But I will post my results for my own record.

First night selections 0 winners from 5 selections.

Profit Loss= -5 units.


	Logged

ATipster

Open
user 1491
Offline

Offline
Alias: None
Posts: 106

2011-May-23, 12:15 PM

21st May - 0 from 5 selections.
22nd May - No Selections

Total Profit / Loss to date : -10 units.


	Logged

Pages: [1] 2 Go Up

« previous next »

Racehorse TALK