spiumA «ww t- c-m» u-m»

J-1 I i

I S « « 15 S 5 12 I» 10 17

Source: Chart created with TradeStation® by Omega Research, Inc.

("O"), there were a lot of down signals, again, mostly of the pull-backs-in-trends type. There was a clear multiple bottom buy in the beginning of October.

The integrated system produced good simulated trading results, as can be seen in Table 18.7. The drawdown ($15,000) is a little higher than what we would prefer, but the system made a 765 percent return on the account in three years ($117,050 net profit); 75 percent of the 164 trades were profitable. The average trade was $713.72, with some trades that were quite a bit more profitable than that (e.g., a trade on 3/1/96-3/4/96 took in $2,450; 3/6/96, exiting the same day, took $2,000). The maximum number of consecutive wins was 17, with an average of two days in the market. The maximum number of consecutive losers was 2, with an average time in the market of 4 days; it only had one loss between 3/1/96 and 5/2/96. It is also a fairly active trading system (e.g., it took five trades in April 1996).

The integrated system performed well over both the longs and shorts: $77,350 profit on the longs, $39,700 profit on the shorts. The drawdown was roughly the same on both and both were highly profitable: 78 percent wins on the longs, 71 percent on the shorts.

Figure 18.7 Signals of a system using all six patterns.

Performance Summary: All Trades

On its own, the integrated system would probably be tradable. This might even be truer if the system were traded with options, as the options would cushion the large drawdown. And, with a profit target, if options were purchased that were somewhat in the money, it would not weaken the profits too greatly due to premium expansion.

Why Does the KCAT Approach work?

Why the KCAT approach works so reliably is a complicated question that does not have any definitive answer. One issue might be synchronization. Consider the analogy between the rather crude method of trying to predict price change over the next two or three bars and an attempt to trace a signal on an oscilloscope. If you do not have a fixed point on an oscilloscope, at which you start tracing the signal each time, you will not be able to see the signal clearly: you will have a jumble of superimposed traces; however, if you always start the scan at a certain point in the signal curve, each time it scanned the curve could appear in roughly the same place. When you are asking the network to predict the change on the next few bars, it is like jiggling the point at which you start the trace on the oscilloscope (i.e., the patterns the network is viewing in the data being presented to it are in every possible position; the network basically has to deal with the patterns being all over the place, not synchronized or located the same way each time. This makes the networks recognition of the pattern a much more complex problem because it is almost as if more patterns are present.

When we use the KCAT approach, by marking the position of the pattern on the chart, we are providing the network with a "synch-lock" point. The network is only being asked to find one pattern that is positioned in a certain way within the

Table 18.7 Trading performance summary using the patterns in Figure 18.7

window of data it is looking at; it is a much easier problem for the net to solve-to find one pattern that is synchronized so that its end occurs at a rightmost part of the window that the network is examining. Rather than asking the network to see an "A" and moving the "A" all around a grid of pixels, we are positioning the "A" so that its rightmost part is always aligned with one line and its topmost part is always aligned with another line; it is a much easier job for the network to recognize an "A" when the "A" is consistently positioned in the same location. Using the KCAT approach, we are always making sure that the network is looking for a pattern that has its rightmost edge at the edge of the window of data that the net is examining.

The KCAT approach only requires the network to recognize a pattern, not to predict; the network is just learning a pattern that existed in the past and being asked to determine whether or not it is currently recognized. The network is not being required to recognize every single pattern that exists in the data; its job is just to recognize one pattern. There are enough variations of the one pattern to still make the task somewhat difficult (e.g., there are all kinds of shapes to pull-backs-in-trends), but at least the net is only having to deal with the multiplicity of subpatterns for that one pattern.

A neural network can, conceivably, deal with recognizing a large number of patterns, but it is then going to need many more neurons in its middle layer, many more inputs, and such a complex net will cause the loss of many degrees of freedom, which means that a hundred times more data will be needed to train the network and achieve consistency. With our method, because one particular pattern is being imposed on the network a priori, the world that the network has to examine is greatly constricted, as is the number of patterns that the network has to deal with. This drastically simplifies the task; imposing such constraint conserves degrees of freedom and lessens the amount of data needed to train the network. This is analogous to doing an experiment where you have an a priori hypothesis and seek either a "yes" or "no" answer versus an experiment wherein you try every possible solution to find the one solution that works; in the latter instance, the result will not necessarily generalize as it could be taking advantage of chance.

These reasons possibly are why neural networks train much better and are much more stable than when we try to use them to develop predictive models based on price change, for example.

CONCLUSION

Through use of the KCAT approach, together with proprietary input preprocessing, we have developed, for both ourselves and our clients, highly successful systems that trade different markets and different time frames, and that can be traded directly on futures without having to cushion drawdowns using options.

The key to using the KCAT approach is in being able to detect by eye, recurrent, recognizable patterns in the market and then being able to consistently mark those patterns on a chart of as long of a time period as possible-a rather labor-intensive task, to