Mar

11

One of the reasons humans are still competitive with computers in chess is that we are aware of patterns that don't compute. Take, for example, nature of pawn structure. One can count individual pawn weaknesses but it's very hard to find an algorithm by which the harmony between pieces and pawns can be assessed. The human mind, however, is quite capable of this.

Might it not be the same with markets, that there are patterns which can't be effectively coded and others which can? As a very simple exercise one might try to count the number of waves that tend to accompany a decline from highs or see whether an n or u formation is being created. Seems to me that it's very, very difficult to do this with numbers; but the human eye is reasonably adept.

The problem of course is that without a clear computable definition of what one is looking for there will be too much that is open to "interpretation," so the results could hardly be relied up. So what is the solution?

I've been thinking about a possible way round this but please excuse me if it is scientifically unsound. What about having generic patterns that contain multiple computable definitions? For example one might have major categories like "panic" or "breakout," but then multiple and detailed definitions of what these are, just to be sure that the computer will recognize them but not for something else. Then when it comes to the stats the generic categories are tested rather than the details.

Just a thought.

Vincent Andres adds:

Another example: it's very easy and fast for a human eye to detect if points are aligned; it's quite a long calculus for a computer.

"So what is the solution?"

This is a deep question. One answer is to stop reasoning/computing with "crisp" sets. With crisp boundaries you have indeed threshold effects that make the reasoning/computing discontinuous and unstable. One way of doing that is using "fuzzy" sets. With fuzzy sets, set limits are no more crisp, but continuous. So working on them is more stable and more continuous. Nice applications are for instance in control.

Fuzzy sets are an interesting tool when it comes to trying to represent knowledge and work with it.

It's not a miraculous tool. Yes it is (or was) a buzzword. You can do the best and the worse with it. And it was done and it is done. Like with neural nets, genetic algorithms, etc, etc. Like with statistics, probability theory, etc. But it's a nice (and very mathematical) topic.

From Steve Leslie:

I like your analogy to visual patterns that don't compute. There are similar parallels in poker.

A computer can give exact statistics of making a hand and pot odds, etc. It can also calculate tendencies with a player. However poker is a game of imperfect information therefore much is subject to interpretation.

Now then:

Crandall Addington is one of the great poker players of all time and a true character. I saw him on TV 25 years ago playing in the old World Series of Poker with a $10,000 buy in. This was when no limit hold-em was essentially an obscure game and $10,000 was a lot of money. He was wearing a Mink Stetson. This was before PETA for sure.

He said that limit poker is a science but no-limit poker is an art.

Limit or structured poker contrary to popular belief contains little bluffing. Most of the hands are played straightforward. There are many multi-way pots and almost all hands go to a showdown.

No-limit hold-em is entirely different. Statistics and straightforward play will only take you so far. It is much more a game of playing the table and the opponents. A feel for the game, understanding its ebb and flow, and evaluating the dynamics of the players are critical. The best no limit poker players know when to be tight, when to turn aggressive, when to bluff, and when to truly gamble. This is where experience is essential.

Similarities occur in trading stocks and futures.

There are the fundamentalists. People like the Buffett of 20 years ago, who was a protégé of Benjamin Graham. Martin Whitman and others. They can be the value players and the grinders. They see big picture things and exploit opportunities but only when the balances are tilted in their favor.

There are certainly the quants, people like Mr. Symonds who obviously have a created a superior mousetrap. But of course, they are neither talking nor sharing what they have found to be successful. There are some others such as D.E.Shaw. Once again they are extremely secretive and are constantly working on their algorithms that identify patterns. Many of the employees are PhD's in computer sciences, mathematics, and music. They are the equivalent of the Rand Institute. Guys and gals who sit in seclusion and are constantly perfecting their own "black box."

Then there are those who trade on a combination of statistics and feel. They tend to be excellent at the "feel of the game" and reading the opponents. The Chair is one of the best of these. One of the finest traders in the world who worked for one of the great traders in Soros. Robert Prechter has had significant success trading off of Elliot Wave patterns.

Then there are the floor traders. They are very intuitive and great readers of the market. They get the first look at where the orders are being placed and who is placing them. In Education of a Speculator, Victor describes in detail how one of Soros's traders would enter the elevator to the floor and the bids would change. It became a game of the cat and the mouse.

In summary: There are opportunities for each of these to profit from the market. As each of the above have demonstrated in their abilities to make money time and time again. It then boils down to what kind of game are you are in and an understanding the rules.

From Bill Egan:

Plotting the data different ways pays off all the time. I earned a US patent because I examined bi-plots of ~50 variables and saw something interesting. Further investigation showed a sensible relationship to the physical mechanism I was interested in modeling, and I quickly built a model that has worked for eight years now.

I always use bi-plots. Once I have a feel for the data and can throw out some variables, I will color points in bi-plots by a third variable. I use this to highlight known extreme values, events, or odd experimental results. It often reveals useful patterns to the careful eye. Histograms of the distribution, data percentiles (percentile function in Matlab), and empirical cdfs are also handy. Multi-modal distributions are often interesting and show up in a histogram.

Software like SpotFire makes this very easy, and includes ways to size and shape data points by other variables (although it isn't cheap to buy). You can certainly do this sort of thing with a bit of work in R or Matlab or S+.

Another trick of the trade is to compute correlations among your variables. You can almost always remove a variable that is r^2 0.9+ with another variable. This will cut down on the amount visualization you need to do.

Further thought from Nigel Davies:

What if the most subtle and powerful engine for pattern recognition and synthesis is in fact the human brain? In this case shouldn't we be training ourselves rather than our computers?

Probably the search for patterns does this anyway but this would seem to be another benefit of the Chair's recommendation to hand count.


Comments

Name

Email

Website

Speak your mind

Archives

Resources & Links

Search