I'm sharing this for those who are interested in the subject. https://deepmind.com/blog/alphago-zero-learning-scratch/
I'm sure many of us are familiar with AlphaGo, the AI program by DeepMind owned by Google, that managed to beat the best human Go players, a feat unimaginable just couple years ago. Now there is a new and improved version just one year later, called AlphaGo Zero, that managed to beat the older version by a margin of 100:0. The paper was published 2 days ago. I read this news with lots of excitement because I have a hunch this is where the future of options trading lies.
Ok - this is a bold statement, I agree and I have no evidence to support that statement, but I'm excited because I do see a path forward for using the same technologies in options trading. However, I'm not envisioning a future scenario of a robot trader trouncing human traders. Instead, I foresee a future (at least in the immediate future) where an AI system will sit side by side a human trader, supporting and augmenting his capabilities. AI will become a tool like an OV or TOS, only far more capable and powerful. When that happens, trading without an AI tool will be like fighting a modern warfare with bows and arrows.
For those interested in the technical details, here's the link to the paper - http://www.nature.com/nature/journal/v550/n7676/full/nature24270.html#access
Specifically I see three technologies that are promising to options trading - Re-inforcement Learning (RL), Deep Neural Network (DNN) and Generative Adversarial Network (GAN). As explained in the article above, AlphaGo started learning to play Go by learning from human. Now AlphaGo Zero is no longer learning from human but rather learning by playing against itself, hundreds of thousands of games a day. That's how it can improve so quickly and more important create new plays that human has never seen before. That to me is exciting.
I see the same potential in options trading, using the same techniques, not to create a automated trading agent (at least not in the beginning) but rather to develop a system that assist in creating trade plans. Imagine an "AlphaOptions" that continuously learns on its own by trading hundreds of thousands of times a day, across thousands of underlyings, testing/discovering all kinds of possible trades. Not only can it help us create trade plans that we could not have imagined ourselves, most important of all, the plans will continuously morph and adapt to market changes. This can resolve a few issues that are severely limiting us today.
Currently, we do manual back testing and manual creation of trade plans. This process is tedious, time consuming and hardly comprehensive. When market changes, our plans stay static for too long. We also rely heavily on our own bias and intuition, thus getting stuck with old ideas. AI, on the other hand, can discover trading ideas beyond our biases and comfort zones. It can correlate multiple underlyings and create trading ideas stringing together multiple symbols, a task that is extremely difficult even for highly experienced options traders. Bear in mind, this approach is not the same as algorithmic backtesting. While algo backtesting can help speed up the back testing process but it will not help discover new trading ideas and will not automatically morph the plans based on changing market conditions.
RL is a fast growing machine learning technology that is at the core of AlphaGo. RL can be used to help us develop a trade plan. Let's say - if we want to develop a butterfly based strategy - if a human does it, he'll have to test out different wing size, how much OTMness, what price to pay, find out what's the optimal greeks for adjustments etc.. If you add up all the permutations that one has to consider, it'll become humanly impossible to do it comprehensively. Human need to use intuition to contain the permutations, but our intuitions will become the limiting factor. On the other hand, RL will discover trading ideas on its own. They less we instruct it what to do, the more creative it will be. It will test out all permutations and then randomly create new ways to trade that might potentially change how we view trading. Too far fetched? I think not - AlphaGo has already created moves in GO that human experts have never imagined and changed forever how the game is played.
DNN - is the another machine learning technology. It is good at pattern recognition. DNN can help recognize patterns and correlations in the market which will take human analysts way too long to figure out. Combining RL and DNN is the one of the hottest pursuit in AI now. It's called deep reinforcement learning. That's what AlphaGo uses and that's the technology I'm eager to explore for options trading. Potentially we can use deep RL to figure out which type of trades are optimal in changing environment, ie think changing skew, changing IV, pricing, bid/ask spreads etc..
Lastly GAN - it's an exciting technology that can generate fake data that mimics real life data. Why ? One thing that ML needs is tons of data to train effectively. We may not have enough. For example - how many blackswans have we seen? Handful. That's not enough to train a network. We need 1000x more. GAN can create fictitious but realistic market data to train our network. GAN is now used actively to create things as well. For example - get AI to re-draw a picture in the same style as Van Gogh or Picasso or create music in the same style as (who ever your favourite musician).
Sorry for the long post, but you probably can sense I'm excited !
Can all these be done ? To be honest, I don't know. I'm just a beginner learner but I think I know enough to see a probable path. This path may lead to dead ends. I can't be sure, but I see this as a challenging but intellectually stimulating path that I'm willing to invest a few years of my time exploring. The worst case is to realize it doesn't work in trading but I will definitely pick up an exciting skill critical for the future.
The reason I'm writing this post is to share what I know but also to invite others who have similar beliefs to collaborate. The vision I'm seeing requires a herculean effort and collaborative efforts from like-minded individuals. I'm still in the learning and discovery process. I won't be ready to start developing the system until a year or two later. Join me in the discovery process if you're interested. For those who are completely new to AI and ML and are interested to get started, I'm happy to share what I know about how to get started. The good news is all the technologies I described above are open sourced and there is a huge amount of learning resources. The only limitation is our time and dedication.
Last but not least, one common question is - "Even if what you said is true, the big boys have 1000x more brains that us. They will wipe us out before we can get started". My response is if that logic is true, then the same must be true for options trading itself. Since we can still be profitable trading options, it shows that if we choose a niche carefully, there are opportunities despite the competitions.
OK - I think I should stop typing !
I'm sure many of us are familiar with AlphaGo, the AI program by DeepMind owned by Google, that managed to beat the best human Go players, a feat unimaginable just couple years ago. Now there is a new and improved version just one year later, called AlphaGo Zero, that managed to beat the older version by a margin of 100:0. The paper was published 2 days ago. I read this news with lots of excitement because I have a hunch this is where the future of options trading lies.
Ok - this is a bold statement, I agree and I have no evidence to support that statement, but I'm excited because I do see a path forward for using the same technologies in options trading. However, I'm not envisioning a future scenario of a robot trader trouncing human traders. Instead, I foresee a future (at least in the immediate future) where an AI system will sit side by side a human trader, supporting and augmenting his capabilities. AI will become a tool like an OV or TOS, only far more capable and powerful. When that happens, trading without an AI tool will be like fighting a modern warfare with bows and arrows.
For those interested in the technical details, here's the link to the paper - http://www.nature.com/nature/journal/v550/n7676/full/nature24270.html#access
Specifically I see three technologies that are promising to options trading - Re-inforcement Learning (RL), Deep Neural Network (DNN) and Generative Adversarial Network (GAN). As explained in the article above, AlphaGo started learning to play Go by learning from human. Now AlphaGo Zero is no longer learning from human but rather learning by playing against itself, hundreds of thousands of games a day. That's how it can improve so quickly and more important create new plays that human has never seen before. That to me is exciting.
I see the same potential in options trading, using the same techniques, not to create a automated trading agent (at least not in the beginning) but rather to develop a system that assist in creating trade plans. Imagine an "AlphaOptions" that continuously learns on its own by trading hundreds of thousands of times a day, across thousands of underlyings, testing/discovering all kinds of possible trades. Not only can it help us create trade plans that we could not have imagined ourselves, most important of all, the plans will continuously morph and adapt to market changes. This can resolve a few issues that are severely limiting us today.
Currently, we do manual back testing and manual creation of trade plans. This process is tedious, time consuming and hardly comprehensive. When market changes, our plans stay static for too long. We also rely heavily on our own bias and intuition, thus getting stuck with old ideas. AI, on the other hand, can discover trading ideas beyond our biases and comfort zones. It can correlate multiple underlyings and create trading ideas stringing together multiple symbols, a task that is extremely difficult even for highly experienced options traders. Bear in mind, this approach is not the same as algorithmic backtesting. While algo backtesting can help speed up the back testing process but it will not help discover new trading ideas and will not automatically morph the plans based on changing market conditions.
RL is a fast growing machine learning technology that is at the core of AlphaGo. RL can be used to help us develop a trade plan. Let's say - if we want to develop a butterfly based strategy - if a human does it, he'll have to test out different wing size, how much OTMness, what price to pay, find out what's the optimal greeks for adjustments etc.. If you add up all the permutations that one has to consider, it'll become humanly impossible to do it comprehensively. Human need to use intuition to contain the permutations, but our intuitions will become the limiting factor. On the other hand, RL will discover trading ideas on its own. They less we instruct it what to do, the more creative it will be. It will test out all permutations and then randomly create new ways to trade that might potentially change how we view trading. Too far fetched? I think not - AlphaGo has already created moves in GO that human experts have never imagined and changed forever how the game is played.
DNN - is the another machine learning technology. It is good at pattern recognition. DNN can help recognize patterns and correlations in the market which will take human analysts way too long to figure out. Combining RL and DNN is the one of the hottest pursuit in AI now. It's called deep reinforcement learning. That's what AlphaGo uses and that's the technology I'm eager to explore for options trading. Potentially we can use deep RL to figure out which type of trades are optimal in changing environment, ie think changing skew, changing IV, pricing, bid/ask spreads etc..
Lastly GAN - it's an exciting technology that can generate fake data that mimics real life data. Why ? One thing that ML needs is tons of data to train effectively. We may not have enough. For example - how many blackswans have we seen? Handful. That's not enough to train a network. We need 1000x more. GAN can create fictitious but realistic market data to train our network. GAN is now used actively to create things as well. For example - get AI to re-draw a picture in the same style as Van Gogh or Picasso or create music in the same style as (who ever your favourite musician).
Sorry for the long post, but you probably can sense I'm excited !
The reason I'm writing this post is to share what I know but also to invite others who have similar beliefs to collaborate. The vision I'm seeing requires a herculean effort and collaborative efforts from like-minded individuals. I'm still in the learning and discovery process. I won't be ready to start developing the system until a year or two later. Join me in the discovery process if you're interested. For those who are completely new to AI and ML and are interested to get started, I'm happy to share what I know about how to get started. The good news is all the technologies I described above are open sourced and there is a huge amount of learning resources. The only limitation is our time and dedication.
Last but not least, one common question is - "Even if what you said is true, the big boys have 1000x more brains that us. They will wipe us out before we can get started". My response is if that logic is true, then the same must be true for options trading itself. Since we can still be profitable trading options, it shows that if we choose a niche carefully, there are opportunities despite the competitions.
OK - I think I should stop typing !