4 ms·

Hey, I didn't actually intend this to be a course. I do not make any money in the market right now so am certainly not qualified to teach a course on it. And of

Hey, I didn't actually intend this to be a course. I do not make any money in the market right now so am certainly not qualified to teach a course on it. And of course, if I was making money in the market I wouldn't have posted this at all. So please everyone remember that. These comments have made me realize it's probably for the best if I do not post the source code. Basically you are competing against armies of PHDs who are buying buildings next to the exchange so they can get their executions slightly faster. It is indeed surprising to me that I was able to make money in the first place. But I do know for a fact that I did make money and I also know that I was not at risk of losing a bunch of money. As mentioned the most I lost in one day was $2000. That's all I was risking.

bjourne 14y ago
Guess you are familiar with http://en.wikipedia.org/wiki/Survivorship_bias http://en.wikipedia.org/wiki/Survivorship_bias? In 2009, there were probably tonnes of people trying to exploit the market using similar low-tech methods as you. Even if all of them were at best break-even, some of them likely made a lot of money on their unprofitable algorithms by pure chance thanks to the size of the cohort. Those few blogged about it and those who lost money didn't. :) I'm not saying that you just were lucky (please dont take this as criticism) - survival bias is just one of those things that always come to mind when people write about how they broke the market or when some investor is presenting his incredibly smart investment strategy that has netted him millions.
- jspaulding 14y ago
  It's a great point and seems like a very smart thing to keep in mind. I think in my case, based on the statistics involved, the odds that my success was luck just seems astronomically small. But, guess I'm biased in my own way :)
  - highfreq 14y ago
    By luck and skill you found a temporary systematic bias that other players missed. It was even luckier that you found it without a lot of upfront losses. But you could have made many attempts and not found any bias and overall lost money and gave up. If lots of people are losing small sums to find these biases, then it may be that the expected return of trying to find biases is zero or negative. At best HFT is a near zero sum game. It isn't creating value for customers. It isn't making the world a better place. It is an unfortunate flaw of our economic system that so many smart people put so much effort into playing zero sum games with each other.
    - femto 14y ago
      An example of your last point. I know a very good engineer, who used to design innovative chips for 4G/LTE mobile telephony. These chips contributed to the market position of one of today's leading mobile phone manufacturers. Today, this engineer is designing ASICs for high frequency trading (basically a specialised Ethernet switch, with all extra logic stripped out, so packets go through a few nanoseconds faster). HFT isn't a zero sum game. It's sucking resources away from productive disciplines into an unproductive discipline, so making a net negative contribution.
      - shrughes 14y ago
        Why are you ignoring HFT's positive contribution?
        damncabbage 14y ago
        Could you please elaborate what that contribution is?
        shrughes 14y ago
        There are plenty of arguments for its contribution. I don't need to repeat them.
        white_devil 14y ago
        It's always the same bullshit excuse: "providing liquidity". It's just that you pretty much need to be another HFT bot to partake in that liquidity.
        lrem 14y ago
        From what I understood, this contribution is not about making stuff nanoseconds faster, but about how this pushes spreads down. Anyone doing any trading will be happier to see the spreads smaller, wouldn't he? Note: by spreads I mean the difference between buy and sell prices. I don't know if there is a special word for it in this context.
  - mcarvin 14y ago
    Please help me understand this better? With a deep understanding of markets and trading I fail to see why you see 'luck' as an explanatory variable is inversely correlated with the frequency of your trades (notwithstanding the effect of trading expenses)? From what I have gleaned the following seems to be true: 1. Your algorithms worked (made money) 2. Then your algorithms did not work, but you could not figure out why If you do not know why something stopped working it seems unlikely that you had a full understanding of why it was working in the first place. Without understanding the nature of the predictive value of the algorithm while it was working, its success seems to be good fortune. Your algorithm could have shown a systematic correlation to any number of factors that could have created strong performance over several months. Performance would then be attributed to accidentally 'timing' a favorable market. I know you feel differently, what am I missing? And either way - kudos on the $500k.
unreal37 14y ago
Thank you for posting this. Very interesting to read. Perhaps posting the source code would not be a good idea, but posting more details would be welcome so that people interested could follow their own path to automated trading.
kolbe 14y ago
FWIW, I couldn't see your pnl chart. I develop algorithmic strategies for a living, and my first reaction to reading your post was skepticism. I'm skeptical for two reasons. (1) because your methods are so unconventional in an industry where convention rules, and (2) because of the time frame of your success, which happened to be one of the more impressive market recoveries in history. I can't tell you how many people I've worked with who fail to isolate the source of their pnl (myself included at times). This is key. It's important to benchmark your strategy against other stupid ones that you know don't have edge. When someone shows me strategies that worked in 2009 and 2010, I immediately make them prove their strategy was not the equivalent of being long equities. Doing this will truly help isolate whether or not luck is involved. When you say that the number and size of your trades justifies the strategy's validity, that's just wrong. You could do 1000 trades in a day: buy 10 RUT futures at the beginning of the day, sell 10 at the end, and just scratch 1 lots for the other 998 trades. In a bull market like 09-10, that would have made 400k, and would have nothing to do with Machine Learning or its applications to HFT. I make all traders benchmark their work against a series of other strategies that I know have no edge, even though they, at times, can appear to have edge. Now, I'm not saying you didn't have legitimate edge, but you do your readers a disservice by omitting relevant stats and discussions like that.
- jspaulding 14y ago
  Thank you. You are right - I should clarify things by saying my program had no directional bias. It was a 50/50 split of longs/shorts. Are there other stats I'm missing?
  - 001sky 14y ago
    vol
  - kolbe 14y ago
    Given a max loss of 2k, we already know the Sharpe Ratio was pretty good. 2009-10 was more than just a huge rally, it was also a period where vol and skew were massively mispriced. I know this is high frequency, but like I alluded to, you need to make sure that what you're doing isn't replicating the pnl profile of low frequency strategies. So, how did you perform relative to vol sellers? From the market bottom to the end of 2010, the max daily loss for a vol seller was about 3x average daily pnl, and >80% winners. So your returns do sound better, but not incredibly. But, even if you failed to perform as well as vol selling did over the same period, that doesn't negate the strategy's validity. If returns were not correlated, then it's safe to say that you weren't just inadvertently shorting vol. So, start there, work out a regression comparing your daily returns to someone selling vol. Do the same with moving average strategies. Mocking up a simple market making back test versus an ES beta is hard, but that too would be a something to test against. I don't expect you to do any of this, and I'm not going to bother to either. I'm just saying that a complete discussion of this subject would include that information.
JoblessWonder 14y ago
I didn't see it in the article and I'm sorry if I missed it... but did you mention what the initial bankroll was?
zygomega 14y ago
I think you undersell yourself - kudos to your success. I'd back a hacker with a plan (and a cash flow crises perhaps) over an army of PhDs any day! Maybe the course you should think about teaching is how to how to orgainse such a high-quality hack as you've described in the article :) I'm currently building a semi-high frequency trading solution and the problem I run into is the sheer breadth of expertise you need to get it all happening. Modern chip design, low-latency, lock-free concurrent messaging, fault-tolerant system design, adaptive learning algorithms, k-means clustering and broker APIs are just a smattering of the ideas I'm trying to get across to make progress. For me, algorithm creation comes more easily than reading about and implementing a broker interface. There is certainly armies of PhDs out there backed by big money but they exist behind heavily guarded intellectual property walls. An open source HFT/Algo/Automated trading platform that brings a hacker sensibility to this problem domain would be seriously competitive.

bjourne 14y ago

Guess you are familiar with http://en.wikipedia.org/wiki/Survivorship_bias http://en.wikipedia.org/wiki/Survivorship_bias? In 2009, there were probably tonnes of people trying to exploit the market using similar low-tech methods as you. Even if all of them were at best break-even, some of them likely made a lot of money on their unprofitable algorithms by pure chance thanks to the size of the cohort. Those few blogged about it and those who lost money didn't. :) I'm not saying that you just were lucky (please dont take this as criticism) - survival bias is just one of those things that always come to mind when people write about how they broke the market or when some investor is presenting his incredibly smart investment strategy that has netted him millions.

jspaulding 14y ago
It's a great point and seems like a very smart thing to keep in mind. I think in my case, based on the statistics involved, the odds that my success was luck just seems astronomically small. But, guess I'm biased in my own way :)
- highfreq 14y ago
  By luck and skill you found a temporary systematic bias that other players missed. It was even luckier that you found it without a lot of upfront losses. But you could have made many attempts and not found any bias and overall lost money and gave up. If lots of people are losing small sums to find these biases, then it may be that the expected return of trying to find biases is zero or negative. At best HFT is a near zero sum game. It isn't creating value for customers. It isn't making the world a better place. It is an unfortunate flaw of our economic system that so many smart people put so much effort into playing zero sum games with each other.
  - femto 14y ago
    An example of your last point. I know a very good engineer, who used to design innovative chips for 4G/LTE mobile telephony. These chips contributed to the market position of one of today's leading mobile phone manufacturers. Today, this engineer is designing ASICs for high frequency trading (basically a specialised Ethernet switch, with all extra logic stripped out, so packets go through a few nanoseconds faster). HFT isn't a zero sum game. It's sucking resources away from productive disciplines into an unproductive discipline, so making a net negative contribution.
    - shrughes 14y ago
      Why are you ignoring HFT's positive contribution?
      - damncabbage 14y ago
        Could you please elaborate what that contribution is?
        shrughes 14y ago
        There are plenty of arguments for its contribution. I don't need to repeat them.
        white_devil 14y ago
        It's always the same bullshit excuse: "providing liquidity". It's just that you pretty much need to be another HFT bot to partake in that liquidity.
        lrem 14y ago
        From what I understood, this contribution is not about making stuff nanoseconds faster, but about how this pushes spreads down. Anyone doing any trading will be happier to see the spreads smaller, wouldn't he? Note: by spreads I mean the difference between buy and sell prices. I don't know if there is a special word for it in this context.
- mcarvin 14y ago
  Please help me understand this better? With a deep understanding of markets and trading I fail to see why you see 'luck' as an explanatory variable is inversely correlated with the frequency of your trades (notwithstanding the effect of trading expenses)? From what I have gleaned the following seems to be true: 1. Your algorithms worked (made money) 2. Then your algorithms did not work, but you could not figure out why If you do not know why something stopped working it seems unlikely that you had a full understanding of why it was working in the first place. Without understanding the nature of the predictive value of the algorithm while it was working, its success seems to be good fortune. Your algorithm could have shown a systematic correlation to any number of factors that could have created strong performance over several months. Performance would then be attributed to accidentally 'timing' a favorable market. I know you feel differently, what am I missing? And either way - kudos on the $500k.

unreal37 14y ago

Thank you for posting this. Very interesting to read. Perhaps posting the source code would not be a good idea, but posting more details would be welcome so that people interested could follow their own path to automated trading.

kolbe 14y ago

FWIW, I couldn't see your pnl chart. I develop algorithmic strategies for a living, and my first reaction to reading your post was skepticism. I'm skeptical for two reasons. (1) because your methods are so unconventional in an industry where convention rules, and (2) because of the time frame of your success, which happened to be one of the more impressive market recoveries in history. I can't tell you how many people I've worked with who fail to isolate the source of their pnl (myself included at times). This is key. It's important to benchmark your strategy against other stupid ones that you know don't have edge. When someone shows me strategies that worked in 2009 and 2010, I immediately make them prove their strategy was not the equivalent of being long equities. Doing this will truly help isolate whether or not luck is involved. When you say that the number and size of your trades justifies the strategy's validity, that's just wrong. You could do 1000 trades in a day: buy 10 RUT futures at the beginning of the day, sell 10 at the end, and just scratch 1 lots for the other 998 trades. In a bull market like 09-10, that would have made 400k, and would have nothing to do with Machine Learning or its applications to HFT. I make all traders benchmark their work against a series of other strategies that I know have no edge, even though they, at times, can appear to have edge. Now, I'm not saying you didn't have legitimate edge, but you do your readers a disservice by omitting relevant stats and discussions like that.

jspaulding 14y ago
Thank you. You are right - I should clarify things by saying my program had no directional bias. It was a 50/50 split of longs/shorts. Are there other stats I'm missing?
- 001sky 14y ago
  vol
- kolbe 14y ago
  Given a max loss of 2k, we already know the Sharpe Ratio was pretty good. 2009-10 was more than just a huge rally, it was also a period where vol and skew were massively mispriced. I know this is high frequency, but like I alluded to, you need to make sure that what you're doing isn't replicating the pnl profile of low frequency strategies. So, how did you perform relative to vol sellers? From the market bottom to the end of 2010, the max daily loss for a vol seller was about 3x average daily pnl, and >80% winners. So your returns do sound better, but not incredibly. But, even if you failed to perform as well as vol selling did over the same period, that doesn't negate the strategy's validity. If returns were not correlated, then it's safe to say that you weren't just inadvertently shorting vol. So, start there, work out a regression comparing your daily returns to someone selling vol. Do the same with moving average strategies. Mocking up a simple market making back test versus an ES beta is hard, but that too would be a something to test against. I don't expect you to do any of this, and I'm not going to bother to either. I'm just saying that a complete discussion of this subject would include that information.

JoblessWonder 14y ago

I didn't see it in the article and I'm sorry if I missed it... but did you mention what the initial bankroll was?

zygomega 14y ago

I think you undersell yourself - kudos to your success. I'd back a hacker with a plan (and a cash flow crises perhaps) over an army of PhDs any day! Maybe the course you should think about teaching is how to how to orgainse such a high-quality hack as you've described in the article :) I'm currently building a semi-high frequency trading solution and the problem I run into is the sheer breadth of expertise you need to get it all happening. Modern chip design, low-latency, lock-free concurrent messaging, fault-tolerant system design, adaptive learning algorithms, k-means clustering and broker APIs are just a smattering of the ideas I'm trying to get across to make progress. For me, algorithm creation comes more easily than reading about and implementing a broker interface. There is certainly armies of PhDs out there backed by big money but they exist behind heavily guarded intellectual property walls. An open source HFT/Algo/Automated trading platform that brings a hacker sensibility to this problem domain would be seriously competitive.