How to define 'its working'

[Xeres]

Trading will be a fifth wheel, leave it as it is.

Are you speaking from experience?

[Xeres]

.... and when you've proved something by backtesting, you'll find that going live with them is a whole other world of pain.

Yes this is very true, I was careful in my modelling, I was creating each day as if it was the actual day, doing the predictions, then moving to the next day and so on but there were still anomilies. Mainly with non-runners, I knew who they were but I didn't know what time in the day they were declared, this impacted several variables in my model, as well as the fact I'm not doing intra-day modelling either. It's been live now for over 14k races in the UK and Ire but I still feel like I could do with 2 full years so I can get a better read on any seasonal impacts.
It's good fun and keeps me busy in the evenings!

[PeterLe]

..Just dashing out but thought Id post a link to a presentation by Noah Silverman

https://www.youtube.com/watch?v=5gW0PO7g6pY

In his summing, he mentions a 'portfolio' ie spreading his bets across multiple combinations. IMO this is good advice
Anyway must dash

(Xeres, Good luck with this; I cant think of a better combination skill set of Maths/programming!

[Xeres]

..Just dashing out but thought Id post a link to a presentation by Noah Silverman

https://www.youtube.com/watch?v=5gW0PO7g6pY

In his summing, he mentions a 'portfolio' ie spreading his bets across multiple combinations. IMO this is good advice
Anyway must dash

(Xeres, Good luck with this; I cant think of a better combination skill set of Maths/programming!

This is the video what set me on my journey, I got talking to a guy a work who made a ton of cash in the early days of Betfair, basically exploiting some loophole in the system. He offered to help me put some strategies in place so off I went. That was the first video I saw on it and went on to develop a model using conditional logistic regression.

The model is specified was more technical than mine, i didn't use any regularisation techniques, or at least I couldn't get them to work as well as just doing the clogit model.

I have since recoded in neural networks because it handles the interaction terms much better, as he sayd trying to code them all up would mean thousands of variables. I use R for all my coding so I am limited to what it can do and I can only afford a modest sized VPS so again I am constrained in those ways.


Yes optimising over many types of bets is something I did think of, I usually pick 4/5 races a day where I get 1st and 2nd place for example. It's just a question of time I guess, I often thought about seeking help but literally nobody I know has any interest in this field. Nobody want to spend their time coding up models in the evening - why not???
Betting is such a dirty word though, as soon as I mention it to anyone they think I have a gambling problem, tradin on the other hand seems to spark more interest.

[napshnap]

Trading will be a fifth wheel, leave it as it is.

Are you speaking from experience?

Yes, from my experience, of course. I'm not good at defining ideal closing positions so I prefer to leave bets unclosed. Modified most of my bots from trading to pure betting and now they generate more commissions and more profit, can't be hurt by outages and by chasing possibility to close my position in illiquid markets. But this's not fit for a scalpers, of course.

So, if you came from gambling background, have a solid and consistent edge and can automate it - there is no point in trading. Just make bets, collect your winnings and be happy.

Betting is such a dirty word though, as soon as I mention it to anyone they think I have a gambling problem, tradin on the other hand seems to spark more interest.

Yeah? Then just say you have a superduper bot and it earns you a few tons per month, and then enjoy the reaction of those incompetent amateurs.!

[pythonic]

I have since recoded in neural networks because it handles the interaction terms much better, as he sayd trying to code them all up would mean thousands of variables. I use R for all my coding so I am limited to what it can do and I can only afford a modest sized VPS so again I am constrained in those ways.

R is convenient for statistics but runs very slow.
You might want to profile your code and do the number crunching in another language.

[Xeres]

I have since recoded in neural networks because it handles the interaction terms much better, as he sayd trying to code them all up would mean thousands of variables. I use R for all my coding so I am limited to what it can do and I can only afford a modest sized VPS so again I am constrained in those ways.

R is convenient for statistics but runs very slow.
You might want to profile your code and do the number crunching in another language.

Doing anything in real time is a bit of a struggle but I have an API that I just pass data to for the predictions and that part is very fast. The training of the models is painfully slow I agree. Tuning my neural net was so slow at one point I had to leave it running overnight.

[Kai]

..Just dashing out but thought Id post a link to a presentation by Noah Silverman
https://www.youtube.com/watch?v=5gW0PO7g6pY

Thanks, it seems like an interesting presentation, just wish there was an English version of it.