and finally analyzed the results, there has been a substantial investment made. Hard work has resulted in provmg one of four things:

1. The sjstem is profitable.

2. The sjstem is not profitable.

3. The sjstem cannot be implemented effectively.

4. The whole process is too expensive.

If youve gotten to the end and your result is not satisfactory, you might want to reexamine the process that you used. Trading strategies should be built slowly; each piece must be proved to be sound before you go on to the next step. You should never have gotten to the end without knowing that the final product was going to be profitable. If the performance was deteriorating as you added new rules and features, you have stopped and reconsidered these features and modified each step to be more robust.

You may also find that your rules were too subjective, that you substituted a rule that was not well thought out, because you were not able to clearly define one part of your strategj. Knowing that you have been subjective about a trading decision is a valuable discovery. Many speculators believe that they are very clinical about their trading policies; they know exactly what will be done in every circumstance and, therefore, think that their approadi can be well defined in writing. But computerization cannot be done successfully if there are any "maybes," ",wait and see," oi any other exceptions.

one common problem is that the sjstem does not work using either the rules or the variables imposed by the program, if moving average periods ranging from 20 through 50 were first used unsuccessfully, it is likely that you would test intervals of less than 20. if stop-loss points were too big or too small (judging by the size of frequency of the losses), youll try other sizes. Sjstems do not alwajs work as expected and perhaps only a slight change to the rules is necessarj. Negative results are important feedback. They can help create a better strategj before it is too late. After all, if the answers were known in advance, this would all be unnecessaij. If the results are good, but different from whal was expected, you should take the time to understand why Dont accept the results without carefiil review. This is an opportunity to either increase your understanding of how the maiket works, or find that you have been making a mistake all along.

A common result is that the sjstem works in some maikets but not in others. This can be the product of usmg fixed values rather than a rule. For example, a stop-loss of 4 tids would make sense for Eurodollars trading at 4>o yield, but not the S&P at 1100. Try to locate the problem and make the rules more adaptable to changing conditions Once this is done, if the problem persists, you must face the issue of robustness.

Rxpectations

In some work, defining your expectations can bias the process and give poor results. For the development of a frading sjstem, you must first have a sound idea before you start. Along with this idea is whether it works best in the short or long term, the frequency of profitable trades, and even the size of the expected profits and losses. By defining your expectations you can apply special rules and techniques that seem consistent with the philosophy of the method These expectations of the profit profile are particularly important to keep you focused throughout the development and testing of the strategj. If you expect to frade one eadi day and capture about 50°o of the trading range, then signals that occur only once eadi week should immediately cause you to look for an error. If profits are too small or losses too big, then you can carefully shidy how the rules apply to a series of prices and understand whether you have omitted some rules, applied the rules incorrectly, or have made a mistake in your thinking. Without setting down clear expectations in advance, you have no benchmaifc to help you through the development process.

A Winner in Disguise

One important twist on performance should not be overlooked: A consistently bad performer can become a consistent winner by doing just the opposite- Finding a profitable frend sjstem in the range of shorter time intervals may seem to be a hopeless effort. On tiie other hand, sjstematic losses mean that the price movement fails to frend, and every buy signal and sell signal is actually an overbought and oversold condition, predictive of price change. This can

happen in maikets that are heavily traded using sjstems.. what appears at first to be a trend change is a tenporary distortion due to massive buy and sell orders.

Consistent losses do not ahvajs mean that the reverse trade will work. The net losses may be less than the total transaction costs; therefore, the opposite action will show even larger losses The successfiil reverse sjstem is one in which the net losses far exceed the transaction costs

Simple or Complex?

It is not necessarj to create or acquire a sjstem that is exceptionally complex to be successfiil. Simple sjstems, such as a dual moving average crossover, can be profitable as well. The difference between the performance of a simple sjstem and a well-developed, complex method is the rid/reward ratio and the magnitude of the equity swings.

Simple sjstems will catch the trends and take as much profit from the big moves as the more sophisticated approaches. It is the other pattems that make distinctions. Added features should reduce the losses and modifj the trading pattems during nonfrending periods. This may be accomplished by identifjing a trading range, using positions of varjing size, or by the inclusion of economic tbta. Sometimes, more complex sjstems are just more complex, without improving results. In that case, use the simple sjstem, but retain larger reserves than normal, just Because II Doesnt Work in Practice Doesnt Mean It Wont Work in Theory

You thought that you followed all the rules by

1 . Precisely defining the sjstem

2. Using enough test data

3. Evaluating performance by forward extrapolation

4. investing enough to survive the worst losing sfreak

5. Trading a diversified portfolio

6. Following every signal

And still you lost all your money... What happened?

Most likely you skipped some part of the well-planned development process. Careful reassessment will usually show that some comers were cut. For example, new traders often only test their ideas in general by apot checking a few selected markets dining years that seem to have interesting or tjpical movement. They assume that these shortcuts do not affect the results of the sjstem. The advantage of computerized testing is that all years can be tested easily-the procedure necessarj to produce a robust sjstem that woiks in most markets.

There is ahvajs the terrptation to force a sjstem to work if it failed because of one or two large losses. An analjst can find something unique about these losing trades and create a rule that eliminates just those iterns. Some traders may just assume that circumstances would have caused them to pass the frade rather than buy or sell in an obviously adverse market. This tjpe of tampering doesnt work and the reason for many of these problems is explained in the next section on price shodi.

Suppose you are not guilty of any of these infractions, that the sjstem was developed without making exceptions, and all the rules were followed. Then, the most common problem is to misjudge the volatility of the markel or some sjstematic change that has occurred. You might see that the program was more profitable in the earlier years, and that the average profit has been declining and the average loss rising. This might be due to increased participation and more market noise. As this occurs, it is necessarj to increase the period of calculation to compensate for the noise For exanple, in the early 1980s, the eneigj markets were very smooth, and a 3-day moving average generated good profits. As volume increased, the market appeared more erratic, and a 10- to 20-day moving average would have been needed. In the early 1990s, the frend apeed would have increase to 30 to 40 dajs, all due to changing market noises When there is higher noise, it takes longer to identifj the frend. Even when you sit still, the market will continue to change.

start Slowly

After paper trading, to begin trading with an exceptionally small portfolio, one that can be lost 10 times over, will not hurt at the beginning. There must be a chance to see if the sjstem performs the way it appeared in testing. This will be the first opportunity to find out the cost of execution. The slippage may be the factor that separates profits from losses. You may find that, during quiet periods, when the martlet is moving sidewajs, there is small slippage but no profits; during fast martlet, when you expect the laigest profits, the slippage also increases The size and frequency of the losses should be checked as well as the equity cycle. When eveijtiiing passes inspection, the investment can be increased slowly. Patience and thoroughness are important ingredients for success.

Once the sjstem has been checked out, it is even safer to tum the day-to-day executions over to someone else, it may be the only certain way to follow the rules precisely. Try to stay detached and objective, but monitor the results carefully to be sure there are no surprises. Even a laige profit may be an early waming of a problem. Large losses often follow laige profits.

PRICE SHOCKS

Price shocks represent the most significant obstacle in the effort to close the gap between test results and actual trading, or expectations and reality. A price shock is, by circumstance, an unexpected event. Exceptionally volatile price moves are caused by actual news that differs from expectations, such as the Fed raising rates by Po when only VA was anticipated; or, it may be a significant, unexpected political event, such as the abduction of Gorbadiev in 1991. The key word to remember is "unexpected."

If an event was expected, then there would be no price change. The market would have alreadj moved to the anticipated level. Therefore, we cannot expect to profit from a price shock by clever planning, but only by chance. We should never assume that we would be on the right side of a laige, unexpected move in more than 50° of those events Unfortunately, when we back-test a frading sjstem using historic data, we tend not to identifj specific price shocks and freat them as normal, predictable events. We choose sjstems that perform best over a set of parameters, without regard to specific trades that may have been the result of shocks. We judge results by higher profits, lower ride or a combination of statistical values. The results chosen as the best performance often have been the greatest beneficiaij of these unpredictable price shocks.

Making Money with Hindsight

We should alreadj be aware of the problems of bad;-testing. While there are no other alternatives for validating a proposed trading strategj, it is necessaij to look at the problems

more carefully to see a solution. No maiket is more evident than crude oil during the Gulf War, discussed in the previous charter (see Figures 22-1, 22-2, and 22-3). From August 1, 1990, through the end of the War in Februarj 1991, oil prices were driven by news. But not all news is a surprise. In the agricultural maikets, a crop freeze in orange juice or coffee is often anticipated by a change in weather. Before a freeze can occur, the terrqaerature must drop. And that low terrqaerature must he sustained to cause damage. This weather change causes processors to protect themselves by bujing forward confracts or futures. In tum, prices move up.

Because Iraq had been moving froops near the Kuwait border, their intentions were not a surprise; however, diplomacy failed. The threat of an oil simply disruption caused oil prices to slowly rise. Many sjstematic fraders and commercials would have been long on August 7,1990, when Iraq invaded Kuwait.

Identifjing Price Shocks

If we cant predict price shocks, and they seriously impact our historic tests, then we must identifj those shods so that we can correct for their effects. In the end, this may mean that we need to look at the individual frades, compare them against a list of price shods of which we are aware, and decide which of the profits were realistic. Even more important, we must decide which risk was understated. There may not be a safe way of changing this final process, il may alwajs be necessaij to manually review the results for credibility. Some analjsts may find it comforting to think that the computer may not be able to do eveijthing on its own!

FIGURE22-1 A far-readiing price shock. A vivid reminder of price shocks occurred on August 16,199 1, and again 3 dajs later when Russian premier Gorbadiev was abducted and