Really? Wow, it is possible that I have been doing computational finance for too long, heh. The paper was Alex Sherstov's masters thesis distilled into a shortened version. Agreed, the reenforced learning agent assumes some knowledge about maching learning algorithmis, but the trend following (TF) and market making (MM) algorithms are very straight forward. There is also some code stubs available on PLAT SOBI, which kinda of like an "ideal simplified teaching version" of any high frequency intraday automated trading system (blackbox). All these are quite different from Wilmott (I browse through Wilmott sometimes), very few traders post on Wilmott, a lot of modelers and analysts focusing on pricing and risk analysis for exotics. Exotics are of little interest to me since that's the realm of investment banks niches. Here is the standard SOBI algorithm, this should be fairly understandable to most people here, right? http://www.cis.upenn.edu/~mkearns/projects/sobi.html
I agree...and now wonder if anyone has coded that SOBI method and given it a real world trial ? I've now have seen the work that has gone into the Penn-Lehman Automated system project managed by Mike Kearns and it's pretty impressive. Preliminary findings with regards to the effectiveness of market-makers strategies is intriquing. However, some of the test results such as the recent contest won by the two PhD's from the London School of Economics indicate more "reality checks" are needed.....as it's doubtful that winning strategy of manipulating the market would work in the long run. There's been no apparent activity with that project for the past 8 months or so....I wonder if Lehman has dropped it's sponsorship of that research ?
For those of you who work with High Frequency systems: I am finding some profitable methods for currencies using limit and stop orders. The problem is that I'm using Trade Station and I can only see the bid and not the ask. How do you all deal with this? It is looking like I need to get both but was hoping I could get away with it at 200-600 minute bar resolution. Any advice that could save me 3 months of effort? Thanks in advance, Adam
Are you talking about TS2000i or TS8.1 ? Big difference. "InsideBid" and "InsideAsk" are available in the latter. For 8.1, see your color.ini file for a complete listing of all server fields.
yes, I see it. Looks like it does not refrence historical..only current bar. I will see how I can work with this. A common order for my code would be: "sellshort next bar at XYZ limit" and my concern is that many fills in TS would not be fills in the market.
Most server fields should maintain their "history" i.e. High, High[1], High[2], etc. Same should be true with InsideBid/Ask. Too bad they don't have VWAP support for Bid/Ask...as the SOBI strategy uses that information. Note: there have been volumes of discussion about the handling of limit orders. Many-a-great backtested system have failed in real time conditions because of lack of alternate logic to accomodate "failed" limit orders....where price reached that level, but the trade was not triggered due to a long order queue.
rufus4000 sir - thks for the links, excellent food for thought! are you running such a system yourself? which asset classes if i may ask?
Yes, I have a set of automated systems trading mostly derivatives (futures and options) with only a little dabble in equities. My concentrate on derivatives because of the built-in leverage, since my own entity is not a hedge fund and I don't trade fixed income, so the reverse-repo (even equity swap) route is out for me to get leverage.
Lehman's original contract was for 3 years, and that was up by last year, but they have since extended it to put it on maintenance, if you will. The core of the PLAT project was actually the PLAT simulator, which has proved to be very valuable as an simulation market for Lehman to test out their in-house algorithms. Back to SOBI for a moment, it is almost too simplistic (but a good starting point as food for thought regardless). Since I have simulator of my own, one day I got bored so I wrote a simple SOBI client and tested it against 4-5 days of tick data. The overall result was that the SOBI client got killed, but it did exhibit a few stable, profit producing periods. The fact that MM agents did well should come as no surprise. Since by definition, MMs would get involved on a lot of volume, and would shave off the spread in 50-60% of the time (70% is a figure quoted in the paper, but actually very close to reality, speaking from experience). So while the "win percentage" of MM agents are low, and the per share profit is very low, it all adds up to decent profits. For instance, in reality, if a system / trader is to be a "market maker" in any product, the old Spear Leeds saying goes that the MM should account for at least *20%* of daily volume.
Thanks Rufus for that update, but how did you know the disposition of the project ? According to Prof. Kearns, HE scaled-back the project himself due to his lack of time available (He began running a quant-based hedge fund sponsored by Lehman). But back to SOBI....the title of it intriques me....as I wonder what a DYNAMIC order book imbalance algorithm would involve ? This would be a DOBI. First thing that comes to my mind would be a frequency distribution of the price deltas for the sample of SOBI "events". A separate distribution for long vs. short positions. Next would be a X-Y plot of price deltas (X) vs. Degree of imbalance (Y). Then generate correlation stats to see if there is any relationship between the two. Also, check out the first derivatives as well....maybe the rate-of-change of the order book from balance-to-imbalance is an important factor. From there, build a trading model that makes bet sizes using the above information.