Table 17.4 Information content analysis

The normalized information content of each buy and sell type technical indicator is shown.

synthesized using subsets E and F; both networks for each Buy/Sell indicator were tested using subsets G and H. This resulted in eight sets of performance results, as shown in Table 17.5. The data shows that predicting Sell indicators (defined by our annual performance target selection and future time period) is somewhat easier than predicting Buy indicators.

Network Optimization

Typically, the default CPM value of 1.00 does not result in the best performing network model. Therefore, we employed an optimization routine that automatically finds the best CPM value for a specific database and set of input variables, using a search optimization algorithm.12

Two statistical networks were synthesized for the buy and sell indicators; each of the four networks was evaluated on two independent data subsets. Rightmost column shows the networks average absolute error for each scenario.

Table 17.6 shows the results of optimizing for CPM. Here, data subset E was used for training, subset G was used to evaluate each network during optimization, and subset H was used for independent testing. In each case, optimization only slightly increases network performance.

Trading Strategy and Results

While the statistical performances of these models show promise, their real benefit can only be demonstrated in an actual trading environment to determine whether they actually make money. Each network was used to process the technical indicators for a number of stocks over the 10-year period. Here, we present the results for AT&T and IBM; the results for other companies are comparable.

Each model-the Buy network and the Sell network-produces simultaneous indicator signals on the continuous range 0.00 < output < 1.00 for each trading day. Both networks can potentially output any value on this range at the same time. We smoothed the outputs with a three-day moving average and defined activation thresholds for each. The result is either a value of "0" (inactive) or "1" (active) for each indicator on each day ("0" when the signal lies below threshold, and "1" when it is above). Once an indicator begins to activate, it typically remain activated for several days.

Table 17.6 CPM optimization results

Optimizing the statistical network only improves results slightly.

Table 17.5 performance of baseline models

Figure 17.8 Buy/sell indicator periods for AT&T.

2 -,

-2 -

Source: Microsoft Excel.

For our trading strategy, we defined the following rules:

• If no signal is currently present and either the Buy or Sell indicator activates, take the appropriate action.

• Only take action (Buy/Sell) on the first day the indicator signal becomes active; ignore subsequent identical signals until there is a change.

• If one signal is present and the other activates, ignore the second signal while both are on.

• If both signals are present and one deactivates, then take the action of the remaining signal.

• "No Signal" gaps of one days duration are ignored, that is, two "strings" of Buy indicators with a single day of no signal in the middle is treated as a single string of Buy indicators.

In addition, because we allow both long and short positions, we include the following rules:

• Long positions are closed out by a Sell signal.

• Short positions are closed out by a Buy signal, and a new long position is also established at the same Buy price.

Finally, each position was closed out at the end of the 10-year period. Note that there are many methods to interpret the indicators produced by this approach by establishing different sets of rules. Here, we merely present one of many.