Intuition behind my thesis (I hope no one is put off by the fact that it's CS work for Stock Markets - you don't need to know much in either of these areas to understand it - I hope!):
'Day Traders' (stock traders who trade many times a day - high frequency trades that are based on local movement of stock prices and not long term profitability of forecasts of stock) are divided on how they think about trading. Some say that the market is efficient (ie, there is no predictability to take advantage of). They are believers of the 'Efficient Market Hypothesis'. Technical traders employ technical trading rules that are intuitive rules based on observation of recent trends. For example, the moving averages strategy uses two distinct windows of time to look at. One over a large time scale (say the last four hours) and one over a shorter time scale (say over the last hour). Comparing the patterns (upward, downward or flat) of movement of prices within the windows... you decide whether the trend will change or not. This is just an example of how these rules work. So you have many such simple rules based on recent trends etc.
Now when traders use these rules, they use it in their mind (adding or dropping rules as they please) in an intuitive way. They cannot explain or clearly define the circumstrances under which they use various combinations, which they consider the more important rules - or which they use at any given time.
My solution is to allot weights to them while combining them. Now, let a machine learning algorithm (the genetic algo or enetic program) decide what weights to allot to them. How does it do this? Look at many previous trading days... of various patterns of variation and try various allotments of weights.. and see profitability. These are th simulations and this testing of possible solutions is how the evolutionary algorithms (genetic algo/genetic program) is run.
Now, over training days (previous days), the agent learns what weights to allot to the various technical rules. So when these technical rules now analyze current market, each of them gives a suggestion. You weight these suggestions using the weights just obtained, and look at cumulative suggestion. That is the one used for placing the final order.
I hope that clarifies it. It's not an explanation of the thesis... but if you read this.. and keep in mind these things while reading the thesis, things will seem much clearer (I hope).
The document:
http://www.ece.utexas.edu/~hsubrama/Harish%20Subramanian_files/Thesis.pdf
Cheers.
'Day Traders' (stock traders who trade many times a day - high frequency trades that are based on local movement of stock prices and not long term profitability of forecasts of stock) are divided on how they think about trading. Some say that the market is efficient (ie, there is no predictability to take advantage of). They are believers of the 'Efficient Market Hypothesis'. Technical traders employ technical trading rules that are intuitive rules based on observation of recent trends. For example, the moving averages strategy uses two distinct windows of time to look at. One over a large time scale (say the last four hours) and one over a shorter time scale (say over the last hour). Comparing the patterns (upward, downward or flat) of movement of prices within the windows... you decide whether the trend will change or not. This is just an example of how these rules work. So you have many such simple rules based on recent trends etc.
Now when traders use these rules, they use it in their mind (adding or dropping rules as they please) in an intuitive way. They cannot explain or clearly define the circumstrances under which they use various combinations, which they consider the more important rules - or which they use at any given time.
My solution is to allot weights to them while combining them. Now, let a machine learning algorithm (the genetic algo or enetic program) decide what weights to allot to them. How does it do this? Look at many previous trading days... of various patterns of variation and try various allotments of weights.. and see profitability. These are th simulations and this testing of possible solutions is how the evolutionary algorithms (genetic algo/genetic program) is run.
Now, over training days (previous days), the agent learns what weights to allot to the various technical rules. So when these technical rules now analyze current market, each of them gives a suggestion. You weight these suggestions using the weights just obtained, and look at cumulative suggestion. That is the one used for placing the final order.
I hope that clarifies it. It's not an explanation of the thesis... but if you read this.. and keep in mind these things while reading the thesis, things will seem much clearer (I hope).
The document:
http://www.ece.utexas.edu/~hsubrama/Harish%20Subramanian_files/Thesis.pdf
Cheers.
posted by Harish at 09:42
0 comments
