Strategy Development: Modelling

A place to discuss anything.
Post Reply
User avatar
ShaunWhite
Posts: 9731
Joined: Sat Sep 03, 2016 3:42 am

Thanks, that's useful.

I've got a question which I hope I can explain, the difference seems subtle but it's quite key....

So would you say you are ...
1. Looking for biases/truths/oddities (or however you want to describe them), in the market type in general, in a set of historical data. To arrive at a general conclusion such as X happens Y% of the time when Z condition is true. Perhaps that might be, if horses steam x% in the last y minutes then they drift z% in the final 10s, in handicaps, lefthanded. Or long duration drifts end at overvalued prices. What you'd call with no disrespect 'sweeping statements'

Or
2. Are you modelling to build a picture of the actual mechanics of a that market type. Perhaps something which says 'if 2fav is shortening and wom on fav is 80% then the price of 3fav is likely to drift...blah blah' .....like a Phillips machine if you've ever come across one, they're an Economic Analog Computer. In theory a 'Phillips' style model could be built with no historical data whatsoever, just a knowledge of how the entities interact. The biggest issue with this model is the number of different variables, internal variables such as volume and external ones such as a news feed. Bad weather on a fav that likes it rock hard would affect it's upper drift limit more than one that slightly misbehaves.

https://www.youtube.com/watch?v=rVOhYROKeu4
3983645714_379e726a40_b.jpg
#1 is basically the output from #2 run many times with different external variables and I see #1 would give you a heads-up about what might happen in certain special situations, if you can remember them all ! Absolutely has it's value.

But #2 really is the holy grail for live trading. I'm not sure you can even describe a pre-off horses in a Phillips machine style way without a brain the size of a planet...but maybe that's what we're building subconsciously when we watch 10,000 markets! :idea:


btw Tennis Trader and Soccer Mystic are so close to a Phillips machine style model.
The "Player A breaks" tap opens, "player A price" falls to x and "Player B price" rises to Z. If you add a few taps for "Player A sweatiness" and "Player B limp amount"....it would be damn near perfect :)


Are we saying the Strategy Development has two threads, Backtesting and Modelling, And modelling has two threads, Data Analysis and Physical Modelling?

I've strayed but my main point is about which type of modelling you mean, #1 or #2. I think you get #1 from graft but #2 is the craft.
You do not have the required permissions to view the files attached to this post.
User avatar
northbound
Posts: 737
Joined: Mon Mar 20, 2017 11:22 pm

My current approach is closer to your #1:

- Gather historical DATA

- Come up with MODEL ideas, basically patterns which I feel repeat often enough

- Build algorithms to BACKTEST each model, simulating entry/exit in the market and resulting profit/loss

- Keep models that backtested profitably

- Build BOTS that enter the markets automatically when one of the models' situation occurs
PeterLe
Posts: 3715
Joined: Wed Apr 15, 2009 3:19 pm

If all else fails, a good guess is better than a bad measure :D
If all else fails, a good guess is better than a bad measure

I’ve wrote that twice as I really believe in it

My best strategies have never really stemmed from hours of reviewing data (in truth I’m not that good at it anyway..
My advice would be to try and do someth8ng counter intuitive and buck the trend.
I’m working on something completely new and so different than I’ve ever done before. I have no real evidence to suggest that it will work, but my gut feeling is that it will, so what better way test it than try it?
I just wanted add a different perspective for consideration and throw that into the mix
User avatar
ShaunWhite
Posts: 9731
Joined: Sat Sep 03, 2016 3:42 am

Despite my interest in more formal approaches I'm pretty much in your camp PeterLe. Partly because I don't have a lovely huge set of data and partly because, I can't see me finding anything the boffins haven't seen and exploited. The feeling I've got before on the forum is that this type of edge has a habit of disappearing overnight too.

So I've been focusing much more on live trading and trying to absorb how the machine works in all its glory. My interest in this mysterious 'modeling' lingo comes from feeling I need a way to record and describe what I'm seeing. But perhaps I'm trying to be too clinical and it's best left to the fuzzy logic autopilot?

'Live trading' is pretty much what it says on the tin. What's the name given to trading that's based on models and backtesting etc? Would you call it algorithmic?

Would it true to say that generally, pre start markets suit the live trading approach, ie reactionary, and post start or final result markets benefit from the algo approach? Ie more prescriptive?
User avatar
northbound
Posts: 737
Joined: Mon Mar 20, 2017 11:22 pm

The problem with manual trading, for a beginner like myself, is that I’m not experienced enough YET to trust that the odds will go in a certain direction.

Peter made a brilliant video about this a while ago:
https://m.youtube.com/watch?v=v7VIxbvrmfY

That’s the main reason why I’ve turned to backtesting. To get some ststistical confirmation from past data that, for example, if I see a certain setup 3min before race start, the odds end up in a certain range by the start, 70% of the time.

This would give me the confidence to stick with a trade even though initially goes the other way. But also not panic those 30% of times I make a losing trade, because statistically there’s a good chance I’ll be in profit by the end of the week/month.

So, not modelling and backtesting for the sake of it, but to get a better understanding of the markets.
Bluesky
Posts: 420
Joined: Mon Sep 19, 2016 9:26 pm

ShaunWhite wrote:
Wed Oct 25, 2017 2:23 am
Would it true to say that generally, pre start markets suit the live trading approach, ie reactionary, and post start or final result markets benefit from the algo approach? Ie more prescriptive?
Although I have no statistical proof to back this up. The feedback I have received from traders is that this is what most of them think is the case. I am sure there will be some people who think the opposite and if that makes them money then good luck to them.
sionascaig
Posts: 1053
Joined: Fri Nov 20, 2015 9:38 am

Peters post on the masters is a very good example of building a model to identify potential trading situations: https://www.betangel.com/blog_wp/2017/0 ... s-masters/

I mentioned this approach to a friend and he said the methodology is very similar to the approach used in his longevity research. Maybe there is something to be learned from those academic types after all (or Peter)!

https://www.ed.ac.uk/igmm/news-and-even ... n-lifespan
User avatar
ruthlessimon
Posts: 2094
Joined: Wed Mar 23, 2016 3:54 pm

However, Peter has a nice way of cutting the data.

A golf hole is categorical data.

For someone building a pre-race model, volume, price, time - pretty much all continuous data. This makes model building extremely challenging & easy to curve fit. Race type certainly could be classed as categorical data - however, this alone won't be enough to form meaningful analysis.
spreadbetting
Posts: 3140
Joined: Sun Jan 31, 2010 8:06 pm

northbound wrote:
Wed Oct 25, 2017 8:37 am

That’s the main reason why I’ve turned to backtesting. To get some ststistical confirmation from past data that, for example, if I see a certain setup 3min before race start, the odds end up in a certain range by the start, 70% of the time.
I love the way people seem to assume punters are lined up in an orderly queue to place their bets at set times and use reams of data to try and prove the fact.
User avatar
northbound
Posts: 737
Joined: Mon Mar 20, 2017 11:22 pm

spreadbetting wrote:
Wed Oct 25, 2017 1:36 pm
northbound wrote:
Wed Oct 25, 2017 8:37 am

That’s the main reason why I’ve turned to backtesting. To get some ststistical confirmation from past data that, for example, if I see a certain setup 3min before race start, the odds end up in a certain range by the start, 70% of the time.
I love the way people seem to assume punters are lined up in an orderly queue to place their bets at set times and use reams of data to try and prove the fact.
Mine was clearly an oversimplification
User avatar
ruthlessimon
Posts: 2094
Joined: Wed Mar 23, 2016 3:54 pm

spreadbetting wrote:
Wed Oct 25, 2017 1:36 pm
I love the way people seem to assume punters are lined up in an orderly queue to place their bets at set times
Although modelling how money arrives in a pre-race market could easily prove that statement false.
User avatar
ShaunWhite
Posts: 9731
Joined: Sat Sep 03, 2016 3:42 am

1second price data, for say the top 4 from -10min to post time, for about 6 months is a LOT of data I simply don't have.

I found things in the 200 races I collected but it didn't continue, just not enough data to slice and dice effectively.
User avatar
Orixian
Posts: 78
Joined: Sun Sep 06, 2015 12:36 am

northbound wrote:
Mon Oct 23, 2017 2:58 pm
Interested.

In fact, having started six months ago to trade preplay horses manually every day and NOT having become profitable yet, I’m in the process of building models to learn how certain preplay market setups traded in the past and see if I can find patterns that repeat.

Early signs are encouraging.

I gotta say that I worked as a software developer for 10+ years. Not sure how easy it would be for a non-developer to build models.
Well as a non software developer I can say it is incredibly easy to collect data about markets using betangel as an interface. I taught my self basic VBA earlier this year but I've a data logger that can capture data using nearly all standard excel code (really only needs a few lines of vba). If you didn't want to build your own logger though you could just download them off the forum here there's plenty of great ones about. The hard part is building a data base you can query. I know what I want to look for in the data but the idea of building a VBA data base is quite daunting and I don't know where to begin at present.
User avatar
northbound
Posts: 737
Joined: Mon Mar 20, 2017 11:22 pm

Orixian wrote:
Wed Oct 25, 2017 7:38 pm
Well as a non software developer I can say it is incredibly easy to collect data about markets...
Agree.

In fact I recently coded a (non-Betangel) bot that collects data, but it will take a while to have enough historical data for backtesting. To accelerate the process, there's the option of purchasing 6 months' worth of Betfair historical data, but it's very expensive...
User avatar
ShaunWhite
Posts: 9731
Joined: Sat Sep 03, 2016 3:42 am

Peter, in your video "How to be a profitable trader" https://www.youtube.com/watch?v=JXSMXwaCH44 you discus how you score a market when you're live trading. You mention an "x out of 10" way to rate your confidence. Are there actually 10 things you look for in combination and if so, roughly how long does it take you to assess them? I guess once you know what the 10 things are you can do it pretty quickly. I also wondered if the figure of 10, being nice and round, slightly disguised or simplified what you actually do?

...and do you keep repeating the whole process while you have an open position? That must be tricky as the tendancy is to get tunnel vision on your selection plus just 1 or 2 other things once the chips are down.
Post Reply

Return to “General discussion”