--Mmrtnga\ferage.$2S

CalcaatKi] period 4> E*po(wnlialsmoo1hing,$2Scomm(ss>or

3. Results should be scored in some way to include profitability and risk. Other measurements can be used, bul stability is of great importance.

4. Results, or scores, should be averaged to avoid spikes or unusual distortions.

5. Execution costs, including liquidity factors, must be included in the manner sppropriate to each sjstem. A trend-following sjstem with entries and exits in the direction of the trend should have greater slippage than a countertrend entrjmethod.

6. Make mistakes in the direction of being conservative. If two sjstems have comparable retums, select the one with the lower rid;, if they have similar retums and rid;, choose the one with fewer trades.

Choosing the speed of the trading sjdem is often the most critical decision in the performance profile. Parameters that cause a trend method to trade faster are most desirable, because the higher number of trades adds confidence to the validity of the test results. You might also find that there is more considency in the performance of a faster trading program compared with a slower one that generates the same net profits. To compensate for its advantages, fader trading is more susceptible to execution problems, so that performance will deteriorate quickly if slippage increases. Figure 21-13 shows the change in performance of all test intervals from 10 to 250 dsjs for the moving average method when transaction cods are increased from $25 to $100 per frade. Note that the shortest interval turns from a small profit to a large loss and profits for intervals through 40 dsjs are halved. Some areas of particularly poor performance, such as at 70 and 110 dsjs, are magnified by the increased costs. In general, there is very little impact on the performance of slower parameter selections.

PROFITING FROM THE WORST RESULTS

Under the right conditions, the worst results of an optimization or test sequence may hold valuable information; it presents the worst-case scenario that should be carefully reviewed for sjdem problems, in particulai with regard to rid; confrol. If the area of poor results is broad and continuous, it may also show a way to improve entry and exit timing. Figure 2 I I4a and b shows the continuous results of 1- and 2-variable tests. Both identifj faster trading areas as having maximum loss and slower frends as having best performance.

FIGURE 21-13 Impact of increased transaction cods on performance.

IMM Swiss franc, 1986-1995

calculation penod

-- Moving average, $25 comnnission 4> Movingavetage,S100commission

The worst results are only useful if there is a net profit from taking the opposite position. This is the case only if the net loss is greater than the transaction costs, including commissions and shppage. But even if the worst score has a near-zero rdum before transaction costs, it can be used profitably. Unprofitable results of a fest trend mean that a long or short position taken at that point lasts only a few dsjs and does not indicate a sustained trend. Use the following rules:

1 . Trade a sjstem of two moving averages, a long-term, designated by the best score D2, and the short-term, selected as the worst score D1.

2. When prices cross the long-term moving average, a long position becomes eligible. A long position is entered when the short-term moving average signals a short signal. Positions may be entered on a 1-dsy delay

3. Short positions may be closed out when prices cross above the long-term trendline; they must be closed oui when a long position is entered.

4. Short positions and exits from long positions are the opposite of rules 2 and 3

Because the short-term signals are not good indicators of trends, they can be used as a countertrend entries. Thai allows for better execution when trading Laiger positions or in less liquid markets. The choice of immediate exit or delayed exit is the traders preference. Once a position is held, there is greater risk waiting to exit, yet most exits would be improved by better timing.

RETESTING PROCEDURE

One optimization is never the end of the testing process. If you have successfully finished the development of a sjstem, and have retained the most recent data for out-of-sample verification, then you should retest die sjstem including the out-of-sample data before beginning to trade the program. Any additional data adds robustness to the sjstem. The impact of adding data will depend on the nature of the market during the new period and the amount of (bta originally used in the testing.

As in step-forward testing, some analjsts prefer to test a fixed amount of data, dropping off the oldest. This does not mean that they test only 6 months, or 1 year of data and

FIGURE21-14 Dual use of testmsp. (a) Single-variable test. (b)Two-variable test

retest every month. You may decide that data before a particular date is no longer relevant. In some cases, there are stractural changes that siport ttiat approach, such as the formation of a unified European currency, which would put conttols on the variance in exchange rates between the participating countties even before the inception date. You might also consider a new maifcet with low liquidity during its earlier years as different fran more active, current marfcet conditions. Retesting remains an important way to adjust the sjstem to characteristics of new data. You might decide that retesting is necessary when time representing 5" to 10° of the test period has elapsed. Using the 2-variable test as an ex ample, a shift may be eepected in the area of best performance as shown in Figure 21-15.

Because only a small amount of data was added, the size of the shift in Figure 21-15 should be small. If a large shift actually occurs, then two possibilities should be considered:

1. The data was unusually volatile and inttoduced pattems not previously seen in the data. In this case, the shift in parameters is justified.

2. The data period for the original tests was short and did not include enough pattems to make the parameter choice robust. With a small amount of test data, there is a

FIGURE21-15 Consecutive tests.