Neural Nets

The sport of kings.
Post Reply
RicHep365
Posts: 105
Joined: Thu Nov 23, 2017 9:42 am

Has anybody on here tried to use any machine or deep learning to predict horse price movements? I know it's not particularly successful on stocks, but for a market which has to balance to an overround close to 100, there must be a practical use for them in horse markets. I've used them to analyse results and cut out losing trades placed by algos, but never to try and predict price movements live.
User avatar
marksmeets302
Posts: 527
Joined: Thu Dec 10, 2009 4:37 pm

I tried, and failed. This was a couple years ago though, and I understand these things a little better now. Perhaps a new attempt would give some results, but I'm too busy with other stuff. You'll have to invest some serious amount of time!

And as my friend xitian has pointed out, given the relatively small amount of data available you would probably be better off trying other machine learning approaches like decision trees etc.
User avatar
Euler
Posts: 24675
Joined: Wed Nov 10, 2010 1:39 pm
Location: Bet Angel HQ

I've constantly revisited this during my trading career, but have much more success modelling the market and using that to anticipate price action.
LinusP
Posts: 1869
Joined: Mon Jul 02, 2012 10:45 pm

I have been working on getting up to speed with ML recently:

viewtopic.php?f=5&t=15320&p=138539

I know others who are making considerable amounts from applying ML at a HFT frequency on betfair.
RicHep365
Posts: 105
Joined: Thu Nov 23, 2017 9:42 am

LinusP wrote:
Tue Jan 02, 2018 3:49 pm
I have been working on getting up to speed with ML recently:

viewtopic.php?f=5&t=15320&p=138539

I know others who are making considerable amounts from applying ML at a HFT frequency on betfair.
I've been learning about Recurrent Neural nets, which are good at predicting results using a time series of data. I'd be looking for it to predict the SP outcome of a horse based on trading activity up to a certain point. So rather than looking to high frequency trade, I'd be looking to back or lay say 2 or 3 mins prior to the off at a better price than the eventual BFSP.
User avatar
ShaunWhite
Posts: 9731
Joined: Sat Sep 03, 2016 3:42 am

RicHep365 wrote:
Tue Jan 02, 2018 7:42 pm
LinusP wrote:
Tue Jan 02, 2018 3:49 pm
I have been working on getting up to speed with ML recently:

viewtopic.php?f=5&t=15320&p=138539

I know others who are making considerable amounts from applying ML at a HFT frequency on betfair.
I'd be looking for it to predict the SP outcome of a horse based on trading activity up to a certain point.
All this is probably far cleverer than I could imagine, but good luck predicting how animals will behave at 10s to the off by looking at their reflections in the market 1-60mins ago.
RicHep365 wrote:
Tue Jan 02, 2018 7:42 pm
I know it's not particularly successful on stocks, but for a market which has to balance to an overround close to 100,
That's the illusion that makes it appears to be a simpler task, but it makes it far harder. Every horse is completely interrelated unlike stocks where that isn't implicit. If M&S have a bad day it won't affect BP. If joint fav of two plays up it WILL affect the other joint fav. Company news, politics and leaked accounts are all known and can be modelled, but stock price prediction is still incredibly difficult. A horse's reaction to a stray crisp packet or a bookie's need to dump a position isn't something I think you'll find in old data (old as in 2 mins ago).

Wouldn't it be easier to start by working out the SP of a football match with ML rather than the complete mind-f*ck that a horse market is?

I'm not knocking what you're doing but it's a hell of an ambition for an individual to aim to succeed where institutions have failed.
LinusP
Posts: 1869
Joined: Mon Jul 02, 2012 10:45 pm

ShaunWhite wrote:
Wed Jan 03, 2018 1:53 am
All this is probably far cleverer than I could imagine, but good luck predicting how animals will behave at 10s to the off by looking at their reflections in the market 1-60mins ago.
How often does a horse play up? Not very often and is something that is normally priced into the market anyway.
ShaunWhite wrote:
Wed Jan 03, 2018 1:53 am
RicHep365 wrote:
Tue Jan 02, 2018 7:42 pm
I know it's not particularly successful on stocks, but for a market which has to balance to an overround close to 100,
That's the illusion that makes it appears to be a simpler task, but it makes it far harder. Every horse is completely interrelated unlike stocks where that isn't implicit. If M&S have a bad day it won't affect BP. If joint fav of two plays up it WILL affect the other joint fav. Company news, politics and leaked accounts are all known and can be modelled, but stock price prediction is still incredibly difficult. A horse's reaction to a stray crisp packet or a bookie's need to dump a position isn't something I think you'll find in old data (old as in 2 mins ago).

Wouldn't it be easier to start by working out the SP of a football match with ML rather than the complete mind-f*ck that a horse market is?

I'm not knocking what you're doing but it's a hell of an ambition for an individual to aim to succeed where institutions have failed.
When applying ML / an automated strategy you (I) don't really care about profit at the market/day level its all about long term. There is normally only about a 2-3 markets a day where there are two faves at similar odds even less when one horse plays up.

Some of my strategies have terrible days, the sort of numbers which would scare the average joe but I know long term there is an edge and I know that as long as my sizing is correct and market impact is low they will be profitable long term.

I can't find it now but I think it was xitian who posted a great reply in regards to developing edges by looking at data in a different way.
LinusP
Posts: 1869
Joined: Mon Jul 02, 2012 10:45 pm

RicHep365 wrote:
Tue Jan 02, 2018 7:42 pm
LinusP wrote:
Tue Jan 02, 2018 3:49 pm
I have been working on getting up to speed with ML recently:

viewtopic.php?f=5&t=15320&p=138539

I know others who are making considerable amounts from applying ML at a HFT frequency on betfair.
I've been learning about Recurrent Neural nets, which are good at predicting results using a time series of data. I'd be looking for it to predict the SP outcome of a horse based on trading activity up to a certain point. So rather than looking to high frequency trade, I'd be looking to back or lay say 2 or 3 mins prior to the off at a better price than the eventual BFSP.
I had actually been thinking of doing this but instead using classification to predict if a runner would steam or drift, trying to predict the actual SP seems harder / more error prone.
RicHep365
Posts: 105
Joined: Thu Nov 23, 2017 9:42 am

ShaunWhite wrote:
Wed Jan 03, 2018 1:53 am
RicHep365 wrote:
Tue Jan 02, 2018 7:42 pm
LinusP wrote:
Tue Jan 02, 2018 3:49 pm
I have been working on getting up to speed with ML recently:

viewtopic.php?f=5&t=15320&p=138539

I know others who are making considerable amounts from applying ML at a HFT frequency on betfair.
I'd be looking for it to predict the SP outcome of a horse based on trading activity up to a certain point.
All this is probably far cleverer than I could imagine, but good luck predicting how animals will behave at 10s to the off by looking at their reflections in the market 1-60mins ago.
RicHep365 wrote:
Tue Jan 02, 2018 7:42 pm
I know it's not particularly successful on stocks, but for a market which has to balance to an overround close to 100,
That's the illusion that makes it appears to be a simpler task, but it makes it far harder. Every horse is completely interrelated unlike stocks where that isn't implicit. If M&S have a bad day it won't affect BP. If joint fav of two plays up it WILL affect the other joint fav. Company news, politics and leaked accounts are all known and can be modelled, but stock price prediction is still incredibly difficult. A horse's reaction to a stray crisp packet or a bookie's need to dump a position isn't something I think you'll find in old data (old as in 2 mins ago).

Wouldn't it be easier to start by working out the SP of a football match with ML rather than the complete mind-f*ck that a horse market is?

I'm not knocking what you're doing but it's a hell of an ambition for an individual to aim to succeed where institutions have failed.
I already have something that beats the SP that uses no machine learning, it’s fairly basic in the grand scheme of things. The edge is small but big enough to be profitable. So I figure with ML it can be improved upon. Got to be worth a try.
Post Reply

Return to “Trading Horse racing”