Earlier quoted context omitted.
What type of analysis did you look into? i.e. what were your input vectors / output?
I would generate a back or lay output signal for 'Correct score' bets for soccer matches before the game started. Input would be everything you can pull out of the the data set especially including the odds on other more standard bets like the match odds and the over / under odds. There are also arb opportunities between the various arcane bets, but that involves speed more than ML.
Ask HN: Has anyone used machine learning to trade?
31–40 of 64 posts
Re: Ask HN: Has anyone used machine learning to trade?
#32Noticing comments on the high degree of proficiency of the competition(or so is the perception.)
Anyone know of any particular markets where a small timer might have a better go at it due to insufficient volume to attract the bigger players, or for any other factors?
Re: Ask HN: Has anyone used machine learning to trade?
#33Being a "small fish" is actually an advantage since you have less capital you need to put to work and thus can more easily get fills and catch the bigger waves generated by the bigger fish institutions like mutual and hedge funds. The institutions have to hold on for longer periods and can't really "scalp" trades per se given that the bid/ask volume would collapse if they didn't distribute and "average in" slowly ove…
Being a small fish means you get killed by brokerage fees, which don't get cheap until the value of your trades is in the millions.
True. Each side has their own set of advantages/disadvantages. And so you'd have to adapt and take that into consideration in addition to bid/ask spread. It's not impossible (nothing is impossible), and you'd only get killed by brokerage fees if you're overtrading.
If you can find an algorithm that can trade only a few times a day and ride stocks that has a volatility of 1% or more then even at retail commission levels you can still do pretty well as you slowly build up your account to trade strategies that require more churn.
Re: Ask HN: Has anyone used machine learning to trade?
#34Yes, I did work on that. I used a neural network to trade FX. I wrote my own neutral network framework using code generators for x86 and CUDA.
The neural network was a fairly normal feed-forward network with GA-inspired, but heavily modified learning process. It combined layers of neurons with memory and memory-less neurons. It is essential to keep the network size small, because a big network will just memorize the data you use to train it.
Also, I used several other inputs, such as the frequency of trades per minute, overnight rates for each currency, and stock market indexes for the relevant countries.
Finding historic data was the biggest problem. I used Bloomberg data from my university, but that was a pain in so many ways. There is NO free data for most of the inputs I needed. Good data from Thomson Reuters or something like starts at $5k/mo.
Overall, the model worked fairly well, generating about 10% annual returns even with a few handicaps I put to be on the safe side. This was about an year ago and the only reason I stopped working on it was that my partner in the financial startup quit. I'm sorry if some of my terminology is wrong - I haven't been working in that field since then. If you have any questions, I'd be happy to talk about this. My email is in my profile.
Re: Ask HN: Has anyone used machine learning to trade?
#35Can't you just download a historical set and use that to test whatever algorithms you think you would use? Then when you are confident enough in their performance, you can take the risk? I wouldn't say it would be a complete waste of time because you would be sure to learn something. If you think you can realistically make money--well then, that's a different question that I wouldn't be fit to answer.
> Can't you just download a historical set and use that to test whatever algorithms you think you would use?
Be careful - much of the historical data has been pruned to eliminate some data from companies that didn't survive.
I've heard that one of the major quant houses (Shaw?) can often guess what data set you used to test your strategy by looking at the strategy.
Plus, there's a selection or fitting problem. Your strategy may have worked well on average from 1972-1999, but that doesn't mean that it will work today.
And there's a velocity problem. There are pricing errors in old data to exploit. And, if you had today's computers then, you might have been able to find them fast enough because your systems are faster than what folks had then. However, today you're competing with folks who have faster systems than you do.
Re: Ask HN: Has anyone used machine learning to trade?
#36I have friends who do this, and they told me, anecdotally, that for every 10 trades they do, 8 or 9 lose money. While they do make it up on the rest, you have to ask yourself if you can afford the costs (brokerage commissions and fees) of executing your strategy in that context (they all work for banks and hedge funds). If you really want to get into this, you might be better off applying to a quant-based hedge fund.…
"I have friends who do this, and they told me, anecdotally, that for every 10 trades they do, 8 or 9 lose money."
At least an AI I would create to trade would probably not be worse than me.
Re: Ask HN: Has anyone used machine learning to trade?
#37Re: Ask HN: Has anyone used machine learning to trade?
#38Can't you just download a historical set and use that to test whatever algorithms you think you would use? Then when you are confident enough in their performance, you can take the risk? I wouldn't say it would be a complete waste of time because you would be sure to learn something. If you think you can realistically make money--well then, that's a different question that I wouldn't be fit to answer.
Trading with historical data isn't accurate enough. It doesn't account for the way your trades would have influenced the market.
Re: Ask HN: Has anyone used machine learning to trade?
#39Re: Ask HN: Has anyone used machine learning to trade?
#40Earlier quoted context omitted.
Do you think that interday would be more suitable then?
A strategy: run a LOT of simulations, keep running them every day. When they are too good to ignore, go raise some capital.
people run MC all the time, not really a new strategy -- expect no breakthrough on that...