Reply
 
Thread Tools
  #934  
Old Jul 11, 2010 9:01am
redpine's Avatar
Member
 
Member Since Jun 2008
Default

Quote:
Originally Posted by pvpn View Post
In FX (where the book is hidden) I wouldn't be surprised to see falling prices and buyers hitting the offer or rising prices and sellers hitting the bid.
Just curious what you mean by "where the book is hidden". Level II quotes? I actually access Level II and rebuild the books on my machine. I realize it isn't the complete market, but the same is true in equities. So what book are you talking about?

Thanks,

Dan
__________________
.
Reply With Quote
  #936  
Old Jul 11, 2010 5:35pm
redpine's Avatar
Member
 
Member Since Jun 2008
Default

Quote:
Originally Posted by pvpn View Post
Dan, I would be interested if your reconstruction of L2 helps you with explaining some of the price returns you observe and to what %.

My take is that only a small fraction of the orders account for most of the effects on price and I think this fraction might be easily hidden/hard to obtain or distinct from the other orders.

Hence, my approach is that I don't try to discover this "flow" but simply assume it's there from the way price behaves and if confirmed by some models I run on a continuous basis.
I originally started looking at imbalances in book orders about 4 years ago. I ended up going down many other fruitless tracks since them, but now am back to book orders. A single day of EUR/USD Level II quotes consists of about 850,000 entries. That's a lot of data and very hard to analyze.

I'm trying to take "DATA" and look at the market as a series of events delineated by something other than time. I think time frames are two arbitrary. An event may take 1 minute or one hour. To me that seems to be why most candle based systems work sometimes and fail the rest. So now I'm looking for those events that change the state of a market. Kind of trying to look at the market as a finite state machine. Those small fraction of orders which you think account for price movement, are triggers the transition from one state the market is in to another. Defining the states and transitions is what I hope to figure out.

This thread seems to be pretty close to what I'm thinking.

Dan
__________________
.
Reply With Quote
  #938  
Old Jul 12, 2010 5:59am
redpine's Avatar
Member
 
Member Since Jun 2008
Default

Quote:
Originally Posted by pvpn View Post
I do almost the same in equities. My program tracks all visible limit orders action and relates this to price returns and the tape. It works for scalping. The "long tail" of earnings comes from decoding the real intentions behind the orders, and this is the real battlefield.
Interesting. That is one of things I have on my TODO list for looking at Time Sales and Level II data. Try and figure out Market orders, when an order is moved and when an order is canceled. I believe it will give a better clue into the motives of the traders in the market.

Quote:
One thing I have on my list is to make (or use) an algo to automatically discover states and transitions. My guess is that there are too many states for a human to try to identify them and connect them with the proper links.

If I remember correctly there was something similar in DrDobbs some years ago.

I hope this is enough to get you started. I would be very interested in your progress.
Start with the three super states. Fear, greed and uncertainty. Then decompose those. If you can figure out the transitions between those states, it could be a trigger for trading.

Dan
__________________
.
Reply With Quote
  #940  
Old Jul 12, 2010 6:15pm
redpine's Avatar
Member
 
Member Since Jun 2008
Default

Quote:
Originally Posted by mikkom View Post
You might want to read about hidden markov models (which I think are a bit too simplidfied to do the actual trick but they are kind of easy to variate). There is a lot of interesting research made on spam filtering (bayesian filtering can be used too when you decide what attributes you want to study) and context discovery (directional network graphs for example) that is quite relevant to state discovery of "hidden" states of any kind of data.

I personally don't use state data per se but my genetic algos can use it if they want to. Basically...
Looked at genetic algorithms and genetic programs. Kinda stuck on neural networks. See some promise with ANNs used in reinforcement learning.

The most important thing I've learned is that it's how you view/prepare the data. Any idiot, like me can use a neural network and yes I did try feeding them OHLC data...

Anyway I don't think it is so much about the state of the market as the state of the traders. Now that could be an infinite problem, but it should generalize with a few thousand autonomous agents in a sim. At the least it will be a fun idea to play with :-)
__________________
.
Reply With Quote
  #943  
Old Jul 12, 2010 6:45pm
redpine's Avatar
Member
 
Member Since Jun 2008
Default

Quote:
Originally Posted by Craig View Post
I don't know if this has been posted yet...
http://arxiv.org/abs/1003.2981
Markov Models??? Hmmm, isn't that what they use to predict the weather? Which is a good point, I really emphatically believe, the further your predictions go out from the current tick, the less accurate they are. Like the weather, they usually get it right today, sometimes tomorrow and rarely a week latter.

Dan
__________________
.
Reply With Quote
Reply

3 Traders Viewing This Thread (2 are members)
withnail, InvisibleMore than 10 Vouchers
Thread Tools


Similar Threads
Thread Thread Starter Forum Replies Last Post
Using future orderflow, to predict spot prices woo Trading Discussion 16 Jul 2, 2012 3:14pm
Vidas - Orderflow Nritv Trading Journals 1 Feb 10, 2011 1:07pm
Help.. Pending stop order when market price and Open price between broker requirement dewaforex Platform Tech 5 Apr 5, 2010 5:32am