JCapper Message Board

General Discussion
-- When do you go live?

Home	Register Log In

When do you go live?

Windoor
5/31/2011
9:27:15 AM

As a "former" pen and paper guy with a limited database at the moment, when do feel confident in going live (real money) on a UDM that is performing well.

My old standard was 100 consecutive playable races at the same track. This was my minimum requirement to have an interest in what I was researching.

Then I would pull, a random month or two from my records to see how it performed.

If good there, I would throw a 200 bank at it to support a $2 win wager. A wise man once told me, if you can't make a profit from a small wager, you won't with a larger one.

Regards,

Windoor.

jeff
5/31/2011
12:45:17 PM

I've always told anybody who will listen the following:

Base each UDM on concepts that identify areas where the public makes mistakes as they bet the races. I talk about this in Finding An Edge. My goal when I create a UDM isn't to have the UDM "select" horses. (Even if... no strike that… especially if you think the UDM will work because the selection process is really good.) This is the approach taken by 99% of your competition. Quick tangent: I’ve spoken at length with a few ADW operators. They tell me that maybe 5 or 6 user accounts out of every 1000 belong to winning players.

My goal is to create UDMs that identify horses that not only have a chance to win but have certain things in their records that cause both the ML linemaker and the public to "X" out the horse on printed PPs and go looking elsewhere for a bet.

Break your data into at least two distinct completely separate samples: a development sample and a validation sample. The RGN (randomly generated number) field in the starterhistory table is useful for this if you are working in sql mode.

Develop each UDM using data from a development sample only. Tune it as you see fit (knowing beforehand that excessive tuning can introduce backfitting into your model.) If you are using a valid concept for the UDM's foundation, tuning (up to a point) can be a good thing.

It's at this point that a lot of players ask themselves "Should I go live?" And it’s at this point where a lot of players fail (for reasons explained below) when they do go live.

Validate your model. Confront the UDM with fresh races taken from a validation sample. Note that races in the validation sample are not present in the development sample. If you introduced backfitting into your model during the development phase, a really good indicator would be that the UDM’s performance during validation testing is really poor compared to UDM performance vs. races found in the development sample. If this is the case, in my opinion, it’s best to scrap the UDM and go back to the drawing board.

However, if your model is based on a valid concept, chances are that UDM performance during validation testing is:

a. Just a little below performance vs. the development sample. (This is quite normal because you very likely introduced some degree of backfitting while tuning the UDM during the development step above.)

b. Positive overall and still quite promising.

If you can validate the model successfully, then consider going live at this point.

more to come...

-jp

.

Windoor
5/31/2011
2:28:01 PM

A development sample and a validation sample.

Makes perfect sense.

How large of a sample would you think is needed for both.

When I do a break-out per track on my first quarter of 2011 database, (my current development sample) it will not return the 100 races I usually look for on a per track basis.

Looks like I will need at least 1/2 year to get near a hundred playable races and some tracks are no longer running.

I would guess, validation would require at least that many again and maybe twice as much?

Looking forward to your "more to follow"

Regards,

Windoor

jeff
5/31/2011
3:59:46 PM

How to go live:

I’ve also told anybody who will listen that unless you have proven ability (first hand experience) growing a small bankroll into something substantially larger, the worst way to go live is to post up a large sum of money and take your best shot.

Players who try it this way, despite the best of intentions, almost always manage to turn a large bankroll into a small one. (Ask me how I know.)

In my opinion, the right way to do it is to post up a small bankroll – no more than a few hundred dollars. Then treat your bankroll with respect. For example, think of it as the last $300 you’ll ever be able to scrape together.

The following formula applies to ALL monies invested by ALL investors in ALL financial markets:

Profit or Loss = (Capital risked) x (Edge)

Etch the above formula into your brain. Make it a part of who you are. When investors put money at risk in financial markets – and pari-mutuel betting is certainly a financial market - the above formula is what truly determines investor success or failure over the course of time.

When you go live, your goal as a player should be to grow your bankroll over the course of time. You do that by repeatedly investing a tiny percentage of bankroll on every qualifying play when you have an edge (or positive expectancy.)

If you’ve done your R&D properly and you really do have an edge, the above strategy, implemented thousands of times over a significant time horizon, will enable you to grow a small bankroll into a large one.

-jp

.

jeff
5/31/2011
10:00:25 PM

--Quote:

"A development sample and a validation sample.

Makes perfect sense.

How large of a sample would you think is needed for both? "
--End Quote.

In my opinion, there is no one size fits all magic "now it's validated" number for sample size. It really depends on what it is you are modeling.

Samples involving UDM foundations that replicate the big picture in a meaningful way (I'm talking about capturing the abilities of the horses and then combining that with the way the public bets the races) often have little variance from one sample to the next. When you are working with a solid foundation it becomes pretty clear after a few hundred race outcomes (in separate samples) that you have captured something relevant in your model.

With weaker models... those that capture some of what's needed from the big picture - it might take a few thousand race outcomes from separate samples to properly set expectations.

One constant here - When you are working with UDM foundations that show a fair degree of variance from one sample to the next... In my opinion that should be your signal to proceed with caution.

-jp

.

~Edited by: jeff on: 5/31/2011 at: 10:00:25 PM~

Windoor
6/1/2011
7:25:37 PM

All good advice. I'll take it to heart.

--quote:

"I’ve also told anybody who will listen that unless you have proven ability (first hand experience) growing a small bankroll into something substantially larger, the worst way to go live is to post up a large sum of money and take your best shot.

Players who try it this way, despite the best of intentions, almost always manage to turn a large bankroll into a small one. (Ask me how I know.)
"

--end quote

I'll assume you learned the same way I did. I have had more than a few 5K banks disappear over time.

--quote:

"Quick tangent: I’ve spoken at length with a few ADW operators. They tell me that maybe 5 or 6 user accounts out of every 1000 belong to winning players."

--end quote

I would have thought that number would be higher. Especially with the rebates.

Still, I would think if one had something that was profitable over time, and lived near a OTB, they might consider sending those plays over there, and play on a cash basis.

Not that I would do such a thing, we all must pay our fair share of taxes.

Testing new ideas with an online ADW is another matter and makes it easier to keep track of results.

Regards,

Windoor

Ishmael
6/11/2011
2:53:21 PM

I think you have to bet real money to get a real take on anything. Not a big believer on theory. Whether it be racing, stock market investing or whatever. That being said you dont want to fall in love with past results and bet yourself into bankruptcy. But I do think you need to bet your UDMS or at least decide why you are not betting them. To be specific as possible there is no training ground like one involving real dollars and real emotions. I know others think different- and this doesnt make me right- but think you need to bet real money. Otherwise its just academic exercise

endearus
7/11/2011
9:44:55 PM

Can someone point me me how I can use the RGN (randomly generated number) field in the starterhistory table to how I can pull and use samples.

Thanks

Eric

jeff
7/13/2011
6:12:53 PM

Sure. Suppose for the sake of argument you have what you think is a pretty decent UPR and you want to create a new UDM using the top ranked UPR horse as a foundation.

As a best practice I recommend using split samples when creating UDMs. The reason for using split samples has to do with minimizing backfitting in the final UDM.

Designate one sample as a development sample. Designate another as a validation sample. Develop and tune the UDM using data from the development sample only.

Then validate the UDM by seeing how it performs when confronted with races not present in the development sample - in this case the validation sample.

Begin developing the UDM with the following expression:


    SELECT * FROM STARTERHISTORY
    WHERE RANKUPR = 1
    AND RGN < 50

The RGN field for each starter in the table contains a random integer between 1 and 100. Using RGN < 50 has the effect of splitting all of the data in the table into two distinct samples (those where RGN is less than 50 and those where RGN is 50 or higher.)

From here, continue developing the UDM adding new factor constraints to the sql expression based on what you see in the Data Window using breakout data from the development sample...


    SELECT * FROM STARTERHISTORY
    WHERE RANKUPR = 1
    AND FIELDSIZE >= 8
    AND MLOR3 >= 1.2
    AND VALF01 >= 80
    AND PSCORE >= 95 -- etc,
    AND RGN < 50

Once you are done and think you might have something, instead of testing the UDM in live play with hard earned bankroll dollars... validate the UDM by testing it against the other half of the database... the validation sample.

To do that, just edit the very last line in the sql expression...

From:


    AND RGN < 50

To:


    AND RGN >= 50

Was backfitting was introduced into the UDM while you were creating it?

To find out, run the UDM through the Data Window against the validation sample (AND RGN >= 50 instead of AND RGN < 50.)

If results from the validation sample are acceptable (the UDM validates) and you want to activate it for live play... You can do that using the UDM Wizard. (Just be sure to remove the RGN reference from the final UDM Definition.)

-jp

.

~Edited by: jeff on: 7/13/2011 at: 6:12:53 PM~

endearus
7/13/2011
12:03:55 AM

I'm really behind the curve.

Thank you soooooo much.

Eric

Windoor
7/17/2011
10:41:21 AM

I used the 2nd quarter of this year to validate 1st quarter results and went live with a small bank 7/1/11. I have three UDM's that look promising after rejecting many and more.

I love this software.

So far so good. We shall see.

Regards,

Windoor

JCapper Message Board

General Discussion -- When do you go live?

General Discussion
-- When do you go live?