back start next


[start] [1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] [14] [15] [16] [17] [18] [19] [20] [21] [22] [23] [24] [25] [26] [27] [28] [29] [30] [31] [32] [33] [34] [35] [36] [37] [38] [39] [40] [41] [42] [43] [44] [45] [46] [47] [48] [49] [50] [51] [52] [53] [54] [55] [56] [57] [58] [59] [60] [61] [62] [63] [64] [65] [66] [67] [68] [69] [70] [71] [72] [73] [74] [75] [76] [77] [78] [79] [80] [81] [82] [83] [84] [85] [86] [87] [88] [89] [90] [91] [92] [93] [94] [95] [96] [97] [98] [99] [100] [101] [102] [103] [104] [105] [106] [107] [108] [109] [110] [111] [112] [113] [114] [115] [ 116 ] [117] [118] [119] [120] [121] [122] [123] [124] [125] [126] [127] [128] [129] [130] [131] [132] [133] [134] [135] [136] [137] [138] [139] [140] [141] [142] [143] [144] [145] [146] [147] [148] [149] [150]


116

"polished" with a very small learning rate for about 20 runs. Training for the in-sample runs ranged between 0.20 and 0.60; in the out-of-sample runs, they ranged between 0.08 and 0.50. In most cases, there was reasonably good generalization.

After the networks were trained and tested, we entered TradeStation® and wrote the trading rules for each of the six networks, which constituted six systems. The entry rule for all six systems was the same, with the exception of a threshold parameter, which varied somewhat from net to net. Thresholds were adjusted for each system to get a reasonable number of trades from each system. As it turned out, the threshold was not too critical for most of the patterns: If it was too low, there would be more signals than would be expected based on the marking procedure and the profits would drop; if the threshold was set too high, it would not necessarily hurt individual trade performance, there would just be fewer trades (e.g., there might be 100% wins, but only with 10 trades). We, therefore, set the thresholds to a middle range, a value that would provide a number of trades roughly equal to the number of instances of the pattern we had marked on the chart, and would also provide an acceptable profit.

The entry rule was "If the networks output is greater than the threshold, then buy/sell next bar." The exit rule was based on a proprietary, "quasi-parabolic" ratcheting stop: Basically, the stop would start out (for the long side) some multiple of the average range below the entry price, and it would go up with the market, becoming like a moving average pulling toward the market as the market rose; in this way, the stop would rise along with the market to lock in profits. Aside from the threshold parameter, the stop acceleration was the only factor that varied across the systems (in a couple of the short patterns, we needed to alter the acceleration factor to prevent market volatility from causing stops to be hit too quickly).

We combined the six systems into one integrated system. We ran the integrated system with a money management stop of $7,500, and a profit target of $2,000. A money management stop is designed to exit a trading position when the trades net loss exceeds a specified amount, in this case $7,500. A profit target is designed to lock in profits by exiting the trade as soon as the trades net profit reaches a specified amount, in this case $2,000.

A Note about the Figures and Tables

The charts are standard bar charts, with opens and closes marked for each bar. Several kinds of signals are generated by the systems on the out-of-sample data, which extended from 2/26/93 to 5/3/96:

• An arrow pointing upward with a 1 indicates a long entry signal.

• An arrow with a -1, pointing downward, indicates a short entry signal.

• An arrow with a 0 and a horizontal bar at the top or at the bottom indicates an exit from the trade.



Results for the Pull-back-in-trends Pattern

Long Positions

Figure 18.1 is the chart for the long side of the pull-backs-in-trends pattern. An upward trend is occurring in the market during this time period. With the exception of the second entry signal (mid October), all entry signals are very close to the kinds of pull-backs-in-trends we marked by hand using the KCAT approach.

Table 18.1 is the performance summary for this pattern, which produced 84 percent profitable trades. There was a net profit of $62,375; a gross profit of $109,550. The drawdown was $10,175, higher than we would like. The system was in the market for an average of 3 days for winning trades, with a maximum number of 15 consecutive winners; losing trades were in the market for an average of 6 days, with only 1 consecutive losing trade at a time. The average trade was $974.61.

This system could be traded on its own. It took 64 trades in the three-year out-of-sample period, which suggests some degree of statistical stability.

Figure 18.1 the long positions of the pull-backs-in-trends pattern.

, t- -us h-sm» uu

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



Performance Summary: All Trades (Long)

Total net profit

$ 62375.00

Open position P/L

$ 0.00

Gross profit

$109550.00

Gross loss

$-47175.00

Total # of trades

Percent profitable

Number winning trades

Number losing trades

Largest winning trade

$ 2450.00

Largest losing trade

$ -6500.00

Average winning trade

$ 2028.70

Average losing trade

$ -4717.50

Ratio avg win/avg loss

0.43

Avg trade (win & loss)

$ 974.61

Max consec. winners

Max consec. losers

Avg # bars in winners

Avg # bars in losers

Max intraday drawdown

$-10175.00

Profit factor

2.32

Max # contracts held

Account size required

$ 10175.00

Return on account

613%

Short Positions

Figure 18.2 illustrates the short positions for pull-backs-in-trends. These signals are not necessarily exactly where we would have placed them by hand, but they are fairly close and illustrative of the pattern; the first short signal is perhaps the closest to the kind we marked.

Table 18.2 presents the performance summary for this pattern. There were fewer trades taken for the shorts (only 26) than the longs; 81 percent were wins. The drawdown was a little lower than that for the longs: $8,975, which is a little on the high side for a net profit of $20,425. The basic behavior of the two systems was similar; however, we would not feel comfortable trading this system on its own because of the small number of trades taken over the three-year period.

Results for the Multiple-Bottoms and Multiple-Tops Patterns

Bottoms (Long) Positions

Figure 18.3 is the chart for the multiple-bottoms pattern (long positions). This chart clearly illustrates the multiple-bottoms pattern and, in many cases, the network marked the bottoms almost exactly where we would have by hand.

Table 18.3 provides the performance summary for this pattern. This system did very well: 95 percent (or 18 out of the 19 trades taken) were winners. Again, the drawdown ($12,025) is a little high for the net profit ($29,500). The number of trades is smaller than we would like for statistical reasons.

Table 18.1 Trading performance summary using the pattern in Figure 18.1



[start] [1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] [14] [15] [16] [17] [18] [19] [20] [21] [22] [23] [24] [25] [26] [27] [28] [29] [30] [31] [32] [33] [34] [35] [36] [37] [38] [39] [40] [41] [42] [43] [44] [45] [46] [47] [48] [49] [50] [51] [52] [53] [54] [55] [56] [57] [58] [59] [60] [61] [62] [63] [64] [65] [66] [67] [68] [69] [70] [71] [72] [73] [74] [75] [76] [77] [78] [79] [80] [81] [82] [83] [84] [85] [86] [87] [88] [89] [90] [91] [92] [93] [94] [95] [96] [97] [98] [99] [100] [101] [102] [103] [104] [105] [106] [107] [108] [109] [110] [111] [112] [113] [114] [115] [ 116 ] [117] [118] [119] [120] [121] [122] [123] [124] [125] [126] [127] [128] [129] [130] [131] [132] [133] [134] [135] [136] [137] [138] [139] [140] [141] [142] [143] [144] [145] [146] [147] [148] [149] [150]