Financial Forecasting Competition

[tenorio@ecn.purdue.edu]

First I would like to apologize to all and specially to Bill Skaggs for not
placing the appropriate emotional indicators in my message to indicate
tongue-and-cheek statements, such as:  8>o   ;>)  :>)

Without these, the message could seem offensive, in spite of the fact that
the sophistation level of the readers here is very high. My apologies to
all.


Bill further wrote:

 But aren't you a little worried that the company that
does best in the competition, even if only by chance, will
take the results as an official sanction and use them in
advertising?  (Feel free to ignore this question if you think
I've already cost you too much time.)

  Thanks again,
        -- Bill

And Steve wrote:



>Its always been curious to me, tho why you would expect people who have
>successful methods
>who may be making money in the market to reveal
>them publicly?  Could this <sampling bias> account for the negative results?
>
>Steve
>
>
>
>
>Stephen J. Hanson, Ph.D.
>Head, Learning Systems Department
>SIEMENS Research 
>755 College Rd. East
>Princeton, NJ 08540


The panel is considering a number of different metrics to be used, and not
declare the winner on a single metric, which could create a winner by
chance. If all the predictions are, for example, poor, indeed declaring a
winner would be a mute point, but it would be very informative as to the
difficulty of the problem. We walk a fine line there and care must be
taken. I don't know how to solve the problem of biased sampling, except to
give them a non-disclosure entry to offer us a counter example that someone
sucessfully can predict the series. If someone was to claim that they can
do the task after the competition, an interesting question would be: so why
didn't you enter it?

Also, all these points are only valid if we are talking about time series
of tradable instruments. Other financial time series would still carry a
lot of value to its prediction, but less of a flashiness, such as interest
rates, sales etc.


Regarding the point about the set of parallel agents for tradable instruments
in the previous message:

All agents have the same policy but different settings. All know about the
current state of the world.

An agent would:
 
- If the market moves by a percentage P up buy, and sell if the market
moves down by the same percentage.

- If an agent is in a certain position (long or short) and the market goes
against them they would sell at a certain percentage drop <= P. 

- If the market goes in their favor they would liquidate their position
after a minimum move of 3P (percentages always measured from a small to a
large number)

Each agent is the same, but the percentages are very different. If there
are more buyers than sellers, the price would go up a point for each extra
buyer. Similarly to the sell side. 

Imagine a sentiment function that turns the sellers into buyers or
vice-versa. This function goes up (say for being inclined to be a buyer) as
the market moves up, up to a point, and then moves down, as the market may
be perceived to be too expensive. At certain threshold points of this
"market feeling function" decisions are made to buy and sell. This function
has a parabolic shape and is recursive, similar to functions in the
logistic family such as:

x(t+1) = r * x(t) * (a-x(t))

which is known to yield to chaotic behavior. The market is then a composite
(sum of threshold versions) of such functions. The variations on P
incorporate trading styles, information, and time scale differences. The
actual function is more like a sin(x) -pi:pi. 

To make the system more sophisticated, the agents may want to have a third
alternative by going neutral before moving from sell to buy
(hysterisis-like). Some agents (small number) can be made contrarians, by
having the reverse behavior.

Further, the market feeling function may be also a function of time with a
decay term associated with slow moving markets.

I plan to write such a simple simulator and place it on the net. If anyone
is interested in beating me to writting the code, and willing to make it
public, I'll help him/her with the task.


About competitions of this kind:

This is not a new idea at all. Makridakis (Journal of Forecasting) and
others have made several competitions/comparisions among various techniques
(mostly linear and in the financial area). Weigend et al. did the same for
non linear techniques with several types of time series. We will be
learning a lot from their experiences as well.

--ft.
____________________________________________________________________________
________________________________________
___________________________

Manoel Fernando Tenorio
Parallel Distributed Structures Lab
School of Electrical Engineering
Purdue University
W. Lafayette, In 47907

Ph.: 317-494-3482
Fax: 317-494-6440

tenorio at ecn.purdue.edu

============================================================================
=