The Forecasting Systems Letter
Jeffrey Mishlove
Vol. 2, # 3
Saturday, April 19, 2003


Market "Phase Shift" Reflected in Breakdown of Leading Indicators

Last October, I reported on a number of variables that appeared to be viable leading indicators of the S&P500 futures contract.  How have these indicators fared since that time?  Let's take a look.  The single strongest leading indicator was reported on October 24, 2002.  It was the four-period Williams R indicator. This indicator compared the high and the low against the highest high and lowest low of the last n number of days.  The following chart, from BioComp Profit's "Chart It" function shows the profitability of this indicator since October 24 -- trading a single S&P contract:

WinPctR as a leading indicator
Four-period WinPctR as a Leading Indicator of the S&P Futures Contract Since October 24, 2002

Here we can see that this indicator continued to function profitably until about the end of 2002 -- but has functioned at no better than chance levels since that time.

One of the other best leading indicators about which I reported last October was the special BioComp Profit variable called WinPctB(7).  This variable compares prices to the upper and lower Bollinger Bands, calculated over a 7-day period.  The chart below shows how that variable has fared over the past six months:

WinPctB(7) as a leading indicator
Seven-period WinPctB Variable (*-1) as a Leading Indicator of the S&P500 Futures Contract Since Oct 24, 2002

Again, we can see that this variable functioned profitably until about the beginning of 2003.  Thereafter results appear to be at levels no better than chance.

Another leading indicator that looked promising on October 24, 2002, was the simple Sine function of the S&P contract price.  The chart below shows that this variable has made a definite reversal in its predictive value:

Sine of S&P Price as a leading variable
Sine of the S&P Futures Contract Price as a Leading Variable

Another interesting leading variable was the reverse of the S&P contract price change.  On October 24, I reported that this variable attained steady positive equity during the period beginning January 1, 1997.  Since then, however, it has also shown a distinct reversal as demonstrated in the chart below:


S&P Contract Price Changes (*-1)(Binary +/-) As a Leading Indicator

How are we to interpret the inability of these leading indicators to continue their performance in 2003?  Perhaps, they were never valid indicators in the first place -- but simply random discoveries of my own search process.  Perhaps, they will kick-in again later this year.  One might argue that the war in Iraq caused a disruption in the normal market functioning.  Or one could argue that a phase shift in the market took place at about the end of 2002.  Perhaps the most reliable conclusion we can draw is simply that, whatever their cause, the market is subject to shifts of this sort.  Effective forecasting systems, therefore, need to be flexible and to be able to learn from the market as it shifts.  This is one reason why I learn towards neural network software -- as it has the potential to shift with the market.


Breakdown of the Volatility Breakout System Reported Last Year

The concept of a market "phase shift" having occurred toward the end of 2002 receives further support from the data I now have concerning the Volatility Breakout System that I had been reporting on last year.  Between October 17 and October 31, 2002, the system had executed three trades (of the S&P futures contract) for a modest profit of  $4,337.

Since that time, the system has executed seventeen additional trades for an accumulated loss of $27, 360.  The following chart, produced in Ward Systems NeuroShell Trader Professional shows the trades executed by this system:


Trades executed since 10/30/02 by the Volatility Breakout System

This chart is a little hard to read.  The solid red and blue triangles indicate positions that were opened.  The hollow triangles specify price targets for closing each position.  If those targets are not achieved, a new target is set for the following day, until the position is closed out.


"0304" -- A New BioComp Profit Neural "Mesh" System
    
The system that I am presenting below is typical of what one can create using BioComp Profit Professional.  The input variables include a mixture of (a) technical measures of the S&P contract price change, (b) OEX options data, (c) advance-decline data for the NYSE and Nasdaq, (d) data on major foreign stock indices, (e) bond market data, and (f) data on related commodity prices.  A total of 27 different input variables were used.

The neural meshes were trained in the same manner reported in the two previous editions of this Letter, on time-series data starting in January 1999 and running through January 26, 2003.  Profit was programmed to generate 2500 different neural models, saving the best 200 of these for a composite system.  The chart below shows the out-of-sample results, starting January 26, 2003.


Neural Mesh System "0304" Out-of-Sample Test Results

As one can see, this system performed very well, indeed, up until about a month ago.  Since then, it has lost about $10,000 in equity -- but still remains a healthy performer overall.  I would expect that this system still has some life left in it.  But, of course, nothing is assurred.  The image below shows the statistics for this system that I have labelled "0304."


Neural Mesh System "0304" Statistics

In spite of an impressive 78.57% hit rate on trades, this system has produced a drawdown of almost $16,0000 in the past ten weeks, lasting twenty-five bars.  The ratio of equity to maximum drawdown is about 2.  However, the relatively short out-of-sample period makes it especially difficult to draw any conclusions from these statistics.  For my own purposes, a neural network system like this is not tradeable by itself.  I would rather use it as part of a "mega-system" as I have previously described (see October 24, 2002).

Forecasting Systems Letter Archive
Return to ForecastingSystems.com