The Forecasting Systems Letter
Jeffrey Mishlove

---

In my October 20, 2002, Letter, I reported that one of BioComp Profit's special indicators, called WinPctB, was the strongest, single indicator I had found at that time.  The indicator in question computer the fraction that the price value was relative to the Bollinger Band of two standard deviations -- computed in an eight-day window.  I had simply chosen eight days arbitrarily, and came up with a very pleasing result.  And, in this instance, I also multiplied the indicator by -1 to achieve positive equity.  Now, with further testing -- using BioComp Profit Pro's "Chart It" function, I have found that the seven-day window for the indicator mentioned above -- provides an even stronger profit curve.  This is shown below:

If you do not have BioComp Profit, and are attempting to replicate this result, you should be aware of the fact that the "Chart It" function normalizes the indicator between +1 and -1.  The subtleties of this feature are further discussed in the article beneath this one.

I also discovered a few other interesting leading indicator for the S&P 500 contract.  The chart below plots the simple sine of the S&P closing price:

The next chart plots the four period, Williams R.  This indicator compares the high and low against the highest high and the lowest low of the past seven days.  This indicator is now the strongest I have found, to date:

Finally, just for fun, I used BioComp Profit's optimal frequency function -- and discovered that one can actually create a perfect wave that also functions as a leading indicator of market price changes:

The "Chart It" function appears in the "input transformation" window of BioComp Profit Pro.  I have referred to charts produced by this function extensively in previous editions of this Letter.  I have found it to be a very quick and simple way of screening potential input variables.  I have referred to several of these, in earlier editions of this Letter, as "leading indicators" for the S&P 500 contract.  And, I have also pointed out what I called a "phase shift" in some of these indicators.

Further experimentation, with a simple input such as Chg(Close)*-1 has revealed to me a new level of complexity in the "Chart It" function that I had not hitherto appreciated.  Specifically, "Chart It" automatically normalizes the specified indicator between +1 and -1.  Let me explain what effect produces.  The chart below shows the raw data from Chg(Close)*-1 for the S&P contract:

On October 20, 2002, in this Letter I presented a chart showing a distinct "phase shift" in this indicator that occurred approximately in March 2000.  That chart was based on a normalization of the above raw data between +1 and -1.  However, if I "normalize" the data myself, by applying, for example, the Binary +/- function, with the threshold set to 0 -- abracadabra!  The "phase shift" disapears.  The indicator generates positive equity steadily over the period since 1/1/1997.  This is shown below:

Another way to achieve comparable, manual normalization would be to use the hyperbolic tangent function in BioComp Profit.  This is shown below with results that differ only slightly from the chart above:

Perhaps you are wondering, at this point, just how strong a leading indicator must be in order to potentially generate $378,925 in profit, trading a single S&P contract, since January 1, 1997.  Unfortunately, BioComp Profit does not provide an answer to this question.  The "Chart It" function has fewer bells and whistles than Profit provides for a neural network system.  One cannot see the actual trades, nor even the summary statistics.  For that purpose, I created a simple system using NeuroShell Trader that provided me with a range of trading statistics:

This system generated a total of 752 trades since January 1, 1997.  Of them, 398 (or 52.9%) were profitable.  The overall ratio of gross profit/gross loss was 1.29.  The average winning trade profit was $3985.  The average losing trade loss was $3499.  In other words, if it is consistent, an indicator such as this need only have a slight edge on the market to generate significant profits over time.  

The maximum,  open trade drawdown was $42,750.  The maximum, peak-to-trough drawdown was $84,400.  So, even system that produces consistent profits over nearly six years, can be quite a roller coaster ride.

My "mega-system" has been performing so dismally in the past ten days, that you might think I should be embarrassed to even talk about it.  Indeed.  But, if I were that easily embarrassed, I should be in a different business.  Short-term, market timing is never easy.  It is always highly risky.  That said, I am going to share with you how I used Microsoft Excel to combine data and provide me with an overview of many different Profit neural network systems at once.  After downloading the most recent market data, I use Profit's "Batch Out" function to create .csv (spreadsheet) files of the signals and equity from each system.  Once that task is completed, I use an AutoMate macro (written for me by computer consultant Lars Spivock) that causes Excel to gather data from each of the "BatchOut" files that I specify and assemble them into a single spreadsheet.

Excel then computes the profitability of each of my neural network systems (now about 27) for the following time period: 160 days, 80 days, 40 days, 20 days, 10 days, 5 days, 3 days, 1 day.  The chart below shows a portion of my Excel spreadsheet display:

Naturally, I keep track of the true, out-of-sample date for each system.  Many of these systems have been optimized on recent data -- so their overall profit for, say, the last 160 days, is inflated as a result.  Once I have visual access to the chart above, I manually give each neural network system a weighting -- based on their true, out-of-sample profitability.  All systems get an automatic weighting of 1.  Extra weight is given to the top three performers in each time frame category.  These weights are then multiplied by the direction of the neural network signals and added together for the "Mega-System" vote.

Because different weighting are applied each day to the individual systems, it is possible for a signal to change after it has been issued.  This is, in fact, exactly what has occurred for tomorrow's market -- as shown below. 

If you would like such a system for yourself, you may contact programmer Lars Spivock (whose link is provided above).

Because of my travel schedule, I do not expect to post a Letter on Friday.  However, I will endeavor to post one over the weekend.  Because of travel and other business activities, it is probable that new letters will be more intermittent in the coming two weeks.

Nirvana OmniTrader Composite Technical Forecast for the S&P 500 Futures Contract:
Short-term, aggresssive strategy: Down.

NeuroShell Volatility Breakout System, Daily, for S&P Futures Contract: 
Buy Stop Order for Friday at 894.20.

KWIK*POP Daily Trading Signal (As Confirmed by Hourly Chart)
Out of Market