Portfolio Selection for System Development

Discussions about the testing and simulation of mechanical trading systems using historical data and other methods. Trading Blox Customers should post Trading Blox specific questions in the Customer Support forum.
Post Reply
Demon

Portfolio Selection for System Development

Post by Demon »

This is something I've been mulling over in the back of my mind for a little while and would like to hear other people's thoughts on this: When does portfolio selection take place? Would one select a portfolio of markets they intend to trade and then develop their system on this portfolio? or would they use a universe of markets covering all sectors, use this to develop their system on and then finally select the markets they intend to trade from that universe?

I suppose what I'm looking at here is the question of curve fitting. For example, if you chose the first method, why not go as far as to use different parameters for different sectors? where does optimization end and curve fitting begin?
kmulford
Senior Member
Senior Member
Posts: 41
Joined: Mon May 12, 2003 5:59 pm
Location: Setauket, New York

Post by kmulford »

Darran,

I believe that we all bring biases into everything we do. In trading and testing, I bring a bias toward trading interest rates and (for those who get the joke) FX. This bias derives chiefly from a long career, recently chucked, on Wall Street in bonds (mortgage derivatives, actually). So, I have been intent upon compensating for this bias: I determined my testing paradigm as broadly as I believed to be viable. I have deliberately over-chosen a broad selection of instruments to test and, perhaps, trade.

Here is something to try. Why not test a number, say three or four, of futures from each of the major complexes? Why not include some international markets, too? You might stick with the most liquid futures. You might find that some complexes do not have three or four "liquid" instruments. You might decide to test three or four and interpret the simulation results of the complex as applicable to the liquid one or two instruments within the complex. This will help to ensure reduce the likelihood of over-fitting by increasing the number of observations (within the complex of related instruments).

I believe that computers have made it pretty easy to process loads of data through simulations. So, you might as well test as much data (different instruments/complexes) as is reasonable. Then, let the results help you to select the markets to trade. But be careful to make sure your portfolio selection makes sense. If an illiquid instument produces the best performance parameters (you decide these), you might not wish to trade it (slippage).

Then it's time to think in terms of the system's robustness. Does the system you have developed actually makes sense? Is it profligate with parameters? Do you apply the same parameters to all markets or all similar markets (complexes), etc.?

In short, test from a large population down to a smaller, tradeable one. It is not science, but one can employ scientific methods to this truly fun search.

Best,
Ken
Demon

Post by Demon »

Thanks for the reply Ken. Correct me if I have misunderstood, but you would suggest settling on the parameters of your system on a 'universe' of markets BEFORE deciding on which individual markets to trade.
Troutman
Contributor
Contributor
Posts: 4
Joined: Thu Sep 04, 2003 10:29 am
Location: Portland, Oregon, USA
Contact:

Post by Troutman »

In the current two programs that Defender has released, my market selection was made both before and after testing. There were a couple of requirements that would rule out markets to trade regardless of their tradeability:

* the market had to have a small enough contract size that we could effectively trade multiple contracts for an ultra-small trading account (sub 100K) without compromising our maximum risk levels AND without being constantly locked out of the market.

* the market had to have enough volume day to day that even when the program reached its intended maximum level of funding, Defender's volume would still be reasonably easy to process.

* Since we needed to stay domestic for the time being, the market had to be a U.S. market.

This ruled out probably 90% of the markets we could legitimately trade. Volume killed most of them -- if 30-50 contracts will move the market and we have to trade 1,500 when maximally funded... we can't trade that market. Of the rest, most of them were simply too dangerous to trade in a small account, either because of intra or inter day volatility, and/or contract size.

After those tests had narrowed the field, I ran a couple of fast yes/no tests on the markets that were left, to eliminate any that had inconsistent behavior (too much overnight motion, too much congestion, etc). Frankly, there wasn't all that much left to pick from at that point.

As to optimization, we did a tiny bit based not so much on the specific market, but more on the characteristics of the market. I've got some generic behavior baskets that I put markets into: trendy, choppy and sitcom.

Trendy markets spend relatively little time whipsawing. Choppy markets are more likely to move violently for a few weeks at a time than to produce big, steady trends. Sitcom markets, I named after TV sitcoms because in those markets every day is a new episode that probably doesn't have much to do with the one you just saw, even though the players are all the same.

Coming up next at 9:30, "A Very Special" Pork Bellies. Ohhh, THAT one's gonna hurt...

Anyway, even though my trading programs are mechanical, the rules are based on generic observations about market motion. If a market is rarely going to put on a big trend, then it doesn't make sense to keep hoping that one will arrive. On the other hand, in a market that DOES move steadily for three months at a time, it doesn't make sense to chop off a profitable trade just as the market is starting to go.

So I'll use a slightly different approach for different categories of markets. All of my parameters are flexible based on the market's own volatility and behavior, so I don't tweak and tune to individual data after selecting the general behavior of the market. If the method doesn't produce good results in a given market, either I've got the personality wrong or, more likely, my understanding of market motion is more limited than I thought...

To say it a lot faster and less confusingly, I ended up picking the markets best suited to the way I trade, from the small basket of those that were available to me after filtering. For Defender's program that uses multiple markets at a time, I tried to combine markets that have no fundamental correlation, to the extent that it's possible to do that.
Roscoe
Roundtable Knight
Roundtable Knight
Posts: 250
Joined: Sat Jan 24, 2004 2:06 am
Location: Houston TX

Post by Roscoe »

I am experimenting with dynamic portfolio selection if this is of interest. For details refer to my post in viewtopic.php?p=5725#5725
Post Reply