System Development Statistics Question

[wesmip1]

All,

I am going through the 2 billion combinations for the neurals now that I have written the program to download the results.

At the moment I don't have heaps of data downloaded as I am still getting the results but I have enough to start playing around with (just over 1000 races).

Some simple tests reveal that approx 0.5% of the combinations show a profit of 10% or more ( thats roughly 10 million combinations ).

What I want to know is whether there is a way to determine what of this is based on chance.

I intend to rerun the analysis after I have a few more races over the next few weeks....

Any help is appreciated.

Good Luck.

[Chrome Prince]

Wesmip1,

You need to determine where the profit comes from, i.e. outlying dividends.

How many of the maximum dividend make up the profit.

Did the profit come from a series of races in one month or does it show a steady progression each month for years?

It doesn't matter is the profit is haphazard, as long as it is not restricted to a single set of results.

Hope this gives you some ideas.

So in effect you could end up with monthly P/L like this:

Plus
Minus
Minus
Plus
Minus
Plus
Plus
Minus
Plus
etc

If it's like this:

Minus
Minus
Plus
Minus
Minus
Minus
Minus
Minus
Minus
Minus
Minus
Minus
Minus
Plus
Minus
Minus

Start worrying!

[michaelg]

I don't know anything about computers but one of my friends was able to transpose some results (non racing) into a graph to show the increase/decrease percentage of his survey. As Chrome Prince says, look for a progression of profit, this should effectively cancel the luck factor. Can this be transposed to a graph which would save time and also be an accurate chronicle of the results?

I presume you are collecting the race results from a TAB - does the data also show scratchings?

Out of interest, can your info identify metropolitan racing because my records indicate they may differ with non-metro.

[KennyVictor]

You need to determine where the profit comes from, i.e. outlying dividends.

How many of the maximum dividend make up the profit.
Hi Chrome,
Over the weeks you've given us the odd formula to calculate if a system is likely to be sustainable. I've written stuff down twice at least and promptly lost it. Could you rehash them for us please.

Cheers,
KV

[wesmip1]

michael,

I remove scratchings from the race.
I don't distiquich Metro and non-metro yet. When I do include this I will only be able to do the races from the start of Apr as I don't have the tvf3 for any months before this. I have the neurals back to the start of the year. I wiped out my test data and have so far loaded in only just over 1 month of results but should have the rest in by the end of this week.

Chrome,

Thanks for the input. I was doing that way anyway, wasn't sure if there was anything else I could do.

As far as putting it on a graph, just gives me something else to code in which shouldn't be too hard.

Good Luck.

[Chrome Prince]

Hi Chrome,
Over the weeks you've given us the odd formula to calculate if a system is likely to be sustainable. I've written stuff down twice at least and promptly lost it. Could you rehash them for us please.

Cheers,
KV


Hi Kenny,

This is what I use as a basis to start, (it might be easier to give a practical example of one of my real systems):

Selections: 668
Winners: 326
Strike Rate: 48.80%
Profit: $61.10
POT: 9.15%
Average Dividend: $2.24
Minimum Dividend $1.10
Maximum Dividend $2.90

Stage One

Divide the profit by the average win dividend giving 27 or 8.28% of the winners - a decent percentage of the winners make up the profit.

Stage Two

Divide the profit by the maximum win dividend giving 21 or 6.44% of the winners - again acceptable.

Look at the scatter of profit, is it consistent or is it located in one chunk of results, so broken down yearly...

Year 1 +39.60 units
Year 2 -11.20 units
Year 3 + 3.30 units
Year 4 +14.00 units
Year 5 +18.30 units
Year 6 - 8.70 units
Year 7 + 5.80 units (5 months data)

So now we know that there are more winning years than losing years, therefore the system is reasonably stable or consistent. In fact, the large loss in Year 2 is more likely to be a hazard result.

I devised the system in Year 4 (backtested) and kept track of it live since then.

This is the method I use with all my systems, some I just follow on paper if I'm not confident of future viability.
From experience, this method of testing only works on decent chunks of data, you cannot form an opinion reliably on less than 500 races.

Dividing the profit by the average dividend, and then dividing that by the number of winners, the MINIMUM I will accept is 5%.

Dividing the profit by the maximum dividend, and then dividing that by the number of winners, the MINIMUM I will accept is 2.5%.

I am amazed to see systems touted "elsewhere" that just don't hold up.

Factors to recognise a doomed system:

Lack of data - less than 500 races
Profit coming from inconsistent patterns
Profit coming from bolters
Profit coming from a small percentage of the maximum dividend i.e. 1% or less.
Even dividing the profit by the average win dividend looks crook percentage wise.

Note: sometimes you have to be a little flexible in your assessment depending on the type of system, but you still have to stick to the basic philosophy.

Hope this helps.

[KennyVictor]

excellent Chrome, thanks.