Neural confusion.
 

These neural things have really got me going.

Before I say anything else on this subject I wanted to recognize the efforts put in by "michealg". Thanks for all your time .....I remain amazed.

Enyhoo, can one of you bright sparks help me out in understanding the significant differences between the CP and $ entries.

Hi, Twobets.

Unfortunately I cannot help as I am at a loss here and seldom include either algorithym in any of my testing. It would appear that both should be related in some way but often the CP allocates high points to a certain horse whilst $ gives it almost no points.

I've tried to work out what factors the CP (and $) alogorithym are based even though the neural website explains it. However looking at a horse's career (for simplicity, those that have very few race starts) but am baffled at some of their CP points. The same with the $ alogrithym - I've looked at total prizemoney earned, comparison of races that offer larger prizemoney, and other factors - but there seems to be no rhyme or reason and am unable reconcile the $ points allocated. Maybe there is a key or master factor that governs both algorithyms but trying to determine it/them has so far eluded me.

I believe CP is a weight/class rating for it's career and assumed that $ was an average prize money rating. The ratings for a particular runner may be adjusted depending on it's opposition's ratings. Just guessing.

For the last 20 months or so, the horse with the highest CP rating produces the best ROI of all the criteria. The $ rating has the same strike rate but worse ROI.

Prizemoney ratings seem to be overbet, generally, whether it be consciously or not. The horse with the 2nd highest $ rating has a better ROI than the 1st. This may be the same for API - I don't have it in my database - maybe someone else could check.

Hey Chinbok, could you possibly to list the top 5 or 6 criteria in order based on ROI and strike rate? I'm working on something at the moment, and I was thinking of concentrating mainly on "CP", "CF" and "Tim" because they seemed to do well in trials that I did. But they were only very short trials. I would appreciate it if someone who's got a lot more info could tell me which criteria have had the best ROI and SR.

Just out of interest I had a look at the $ algorithym in conjunction with the default settings. If the top selection also had the most points in the $ category (outright, not equal top $), the Place results since last Monday are (I've not looked at last Wednesday as they were metro meetings):

Monday.
5 selections for 4 placegetters paying $5.40

Tuesday.
6 selections for 4 placegetters paying $6.20

Thursday.
8 selections for 7 placegetters paying $9.10

Total of 19 selections for 15 placegetters.
Strike rate of 78%
POT of 14%.

Chinbok, I might have a look at the above system but instead of using the top selection in $, I will see how it fares with the CP algorithym. I'll post today's selections for both methods, if there are any, in another thread. But I won't include Dombeen in the $ method.

Hi Sportz,

Here are the strike rates and returns over the last 12 months for each of the neural algorhythms on their own - betting on top horse.

CP.......23.4%.......89.4%
CF.......21.7%.......83.5%
Tim......16.7%.......84.3%
JA........18.0%.......84.0%
TA.......17.2%.......87.1%
JT........17.8%.......85.2%
BP.......11.9%.......78.7%
WT......14.0%.......79.8%
Crs.......16.2%.......82.9%
D.........16.7%.......85.9%
$.........22.8%.......83.0%
DLR......13.1%......85.1%

Thanks a lot. That's a great help.

Interesting to see that CP outperformed the other categories in both strike rate and return of money. $ almost matched the CP strike strike rate but a disappointing return.

Here's some similar figures over 16months, not including the last 4.


Criteria SR POT
CP 22.1 -12.4
CF 21.1 -16.3
Tim 15.1 -14.7
JA 16.7 -14.2
TA 16 -12.1
JT 15.5 -16
BP 11 -22.9
WT 12.3 -21.5
Crs 13.6 -17.3
D 14.2 -16.2
$ 22.3 -14.9
DLR 10.7 -19.6

Sum(JA,TA,JT) 18.8 -11
Unitab Rating 22.5 -15.5


Seems I lied about CP having the best ROI, TA is slightly better - not the case in KV's data though.

KennyVictor
 

How much data do you have? How do you use the neurals?

When you say betting on the top horse in each category, is this the outright top selection in each category because quite often there is more than one horse with equal top points? In the method I'm currently testing with CP and $, I eliminated any equal-top selections as they did not perform, even though there were only a few selections. If the neurals are accurate, then being equal top does not really give any one horse a distinct advantage over every other horse in the race.

Hi Chinbok,


Only have data since 1/6/05 (not 100% complete but not missing many races) which I only collected a couple of weeks ago. I have it set up so that I can run queries which set a value to any of the neural algorhythms, can set a minimum value to any of them, set a minimum total and a few other things like must have had x number of races, metro / country / provincial, FGDSH going, etc.
I try various combinations to see how the top horse fares and then I bet a little experimental money on the selections and usually watch my balance go down. I've found a few combinations that show up to 20% profit over the last year but they haven't performed over the last fortnight. Only time will tell.
Of course there is a conspiracy theory that the neural guys have a great system for picking winners but they have shares in the TABs and online bookies - so they get you hopeful with winning systems and then change the algorhythms to make you lose and their shares go up. Well, no there wasn't really a theory of that sort but there is now.

KV

HI michaelg

I set the neural multiplier to 5 for maximum accuracy (lower figures get rounded to the nearest whole number of course) then picked the first highest score. YOu're quite right, sometimes they do come out the same and I didn't test for that. If there were 2 the same it would have picked the one with the lowest tab number. Not 100% accurate but good enough to give you an idea of which neurals perform best on their own.
KV

I have seen you all talk about this but i realy have no understanding, where is the best place to read up on it please.

The neural settings I concentrate on are Tim, JT, Crs, Dist, $ and DLR

The settings have produced:

499 Races
126 Winners (S/R 25.3%)
260 Places (S/R 52.1%)

Win POT is 9.9% or 49.35 units

......so I'm happy this system, although it did run into a negative in the middle of May.... so sometimes it pays to persevere.....

I've concentrated on these settings as I think, horses are a creature of repitition and therefore Tim, Crs and Dist are highly weighted and that JT, $ and DLR are also relatively important factors..... not to say they are the be all and end all..... as michaelg has alluded to, subtle variations and outside filters seem to improve results..... however I must note that I have not used any filters in this system - so regardless of track condition/field size/days break/first starters/points assigned yatta yatta they are all inclusive - warts and all !! - I like the turnover factor!

I too would also like to say well done to Michaelg for some excellent contribution to the forum, for it was because of him that I went back to the neurals.... and I am very happy (so far !!)

Good Luck people... (There is my 2 cents worth :D )

Hi KV,

I like your conspricay theory. I've had a few systems go really well and then couldn't believe how it could turn around. I had a few on the go a while back and made alot of paper money but gave most of it back.

I've also started recording the Winner In Six Ratings but only have ~4months. They've got a pretty good strike rate from the top 2 and manage to find some good priced winners. Hopefully I can use them in conjunction with the neurals - need more data though.

Cheers

KV,

I was wondering if you had progressed with your 'intellectual exercise' idea from a while ago.

I got my brother involved in a similar project recently using the neurals plus Unitab rating and Weight carried below the toppy. He wrote some code that could try 10,000 variations in less than a minute and optimise the factors automatically. He ended up with a set of factors that produced a 4% loss on turnover betting in every race over 16months. The factors were optimised for POT though and the results weren't reproducible in the 4 months of test data that I had. The strike rate was exactly the same though. We'll try again, optimising for SR instead of POT and see how it goes. I don't think it will match the -4%.

Cheers

Gratitude
 

Thanks Masters Chinbok & michealg for your help. I am somewhat amazed at the level of interest generated by this subject, and it suggests to me that the neurals must have shone a little beacon of hope into many a forumite's calculations. You wizz kids with neural data on file must surely have some interesting anomolies within your data.

I must admit, all this neural talk has also had me spending alot of time on the neural site. For what its worth, the major thing that struck me was the strike rate of the top 3 or 4 in the CP factor. Just from what I've seen it seems like it would be around the 65-70 % mark, but that was from a random sample so it may not be the norm. What seemed more remarkable was that many winners in the $10 to $20 bracket could be found in the top 4 CP. So if your out there KV, would you be able to run the following query on your results please, if your able to.

Top 4 CP
$10-$20

Also Kenny, have you noticed the strike rates for any factor are above average in certain circumstances, such things like barriers on some courses may have above average reults, especially narrow tight circuits or short straights. Also DLR with different distances.

Sorry Dr Ron, not going to be much help. I'm only set up to monitor top horse at the moment and don't really have the programming time to change it to top 4. If I do I'll put a note in here.
As for the other more esoteric stuff you mentioned I haven't tried those options and suggest that a year's worth of data probably wouldn't give enough data in any one set to produce meaningful results.

Chinbok, had a few ideas after the intellectual excersise thread but as soon as I thought of a strategy the goalposts moved - I found a bunch of other variables that needed factoring in. I find that rather than say five neural figures with a minimum total for example - it's better to perhaps set a minimum for 2 or 3 of those neural categories to be sure the horse has (say) at least reasonable recent form and has earned some prizemoney. I've only been playing with the neurals for a short time and am still trying new things willy nilly at the moment.

KV

Thanks for that Kenny. Do you have any thoughts on whether the actual neurals points or a ranked order of points would prove a more accurate method?
For example a horse with say 80 cp points would probably not have a 20 times better strike rate with thoise on say 4 points? If you get my drift.

Dr Ron,

This is not exactly what you asked for but might be useful. In the attached file is the average neural rating of all winners at different distance ranges. The actual values are irrelevant because they have been adjusted relative to the highest rater being 100. You can see the importance of DLR as distance increases.

Very interesting Chinbok, and I notice a couple of the others move in the opposite direction as distance increases.

Very interesting results Chinbok, in the system I'm currently working on based on the percentages provided by you and Kenny, I had completely shelved the DLR figures, but after seeing your table will have to work them in starting at 1350 m .

Hi Dr Ron,

From the questions you have asked me I see you work in a very different way to what I do but from my observations (and they are a bit limited I admit) the neural factors really can't be taken in isolation. It's like trying to handicap a horse by saying "Yes, it's won at this distance 3 times out of 4 so there is a 75% chance of it winning this time", when it hasn't had a run for 12 months, has never won in the wet and it's starting from barrier 27. Personally I think the DLR is a factor of great importance, used much as MichaelG uses it. If the horse has a certain minimum in this column it probably means it's ready for its next run.
A nice balance of some of the other factors which feel good and logical together (bit of current form, good jockey, whatever) is possibly the way to build up a winning system.
Not really trying to criticise you here Dr Ron, more saying that my table of single neural returns is not a lot of use on its own. Chinbok's looks to have more value.
KV

I realised this morning, while in the shower, that the average ratings of winners I gave previously could be misleading.

Nearly all horses start their preparations over the shorter distances. These runners will have low scores for CF, Tim and DLR which will bring the average of all winners down. So I've added the average ratings for all runners at each distance range so this can be compared to the winners.

Now it appears as there is no correlation between DLR and distance.

It's like the statistic that (about) 75% of winners have run in the last 21 days when 75% of all runners have run in the last 21 days.

Hey guys, I tried to download some neurals from previous meetings so that I wouldn't have to pester you for info using the method mentioned by someone on another thread, (might have been wesmip1 from memory) but kept getting a message about some sort of script error. Would anyone be able to help? thanks

Very interesting stats.
Does any one have any stats on the results for the std. factory setting.

Cheers.

The script error is probably because the whole site link (including the long form of the venue and date) aren't there and it confuses the server. If you hit the neurals button over to the left though I think you'll find it finds the right data.
KV

Over about 12 months. All courses, all track condt. ca 15000 bets
Win:...24.25% Sr, 88.1% ROI
Place:.53.6% Sr, 92.3% ROI
KV

Thanks for that Kenny, by the way, are those last stats you put based on one tab price only, if so its not a bad set of results and you could probably break even by using either best tote top fluc or betfair. pretty good really for a set of ratings that take minimal time and no cost.

p.s just tried it then kenny, works a treat :)

p.p.s is there a list of track codes any on the site?

A 7.7% loss on place betting. If only the TAB did not deduct such a high take-out, and if Austote was to operate on place betting...