The Optimization Paradox By Curtis Faith, an Original Turtle; partner in Galt Capital, LLP; managing member of Acceleration Capital, LLC; and founder of Turtle Trading Software I've noticed something when discussing hypothetical trading results with people that seems to get little discussion but which can really be misleading, something I call the Optimization Paradox. The Optimization Paradox: Proper optimization results in a system that is more likely to perform well in the future, but is less likely to perform as well as the simulation indicates. So optimization improves the performance of the system while decreasing the predictive accuracy of the historical results. I believe that an incomplete understanding of this paradox and its causes has led many to shy away from optimizing systems out of a fear of over-optimizing or curve-fitting a system. However, I contend that proper optimization is always desirable. NOTE: A complete discussion of the complete process of Proper Optimization is beyond the scope of this article. What is a Parameter? A well known commercial system is touted by its developer as a better system because it has only "one parameter". While the developer may have optimized only one parameter, I believe there are, in fact, many parameters. Many constant values used in the system, like 2, or 2%, or 5, etc. are actually parameters that have not been optimized (or perhaps the optimization has been hidden from the purchasers). The system actually has more like five or six actual parameters, even if the developer does not make it clear, or even believe himself, that these are parameters. Consider a simple moving average crossover system: Using a simple 80 day moving average, buy the next open if the price closes over the moving average, sell if it closes below the moving average. The entry stop is 2 ATR below the moving average for longs, 2 ATR above for shorts. Bet size is 2% of total equity. How many parameters are there in this system? Many people would answer one parameter, the number of days in the moving average. I'd answer differently. First, there is nothing magical about the crossing over the price of the moving average. Just because we have decided that the exact price of the moving average is the threshold to buy, this does not mean that one couldn't choose other prices related to the moving average, say 1/2 ATR higher, or 1 ATR higher, etc. Second, in the stop of 2 ATR, the value 2 is a parameter. Also with the bet size of 2%, the 2 is a parameter. There is nothing magical about the 2, one could just as easily use 1%, or 1.5%. So the 2 is just one value of many that could have been used. Each of these placeholders for values are parameters. The Benefits of Optimization Optimization is always beneficial when done correctly while accompanied by a mature understanding of its implications. The basic reason is that it is always better to understand the performance characteristics of a parameter than to be ignorant of them. Optimization is simply the process of discovering the impact on the results of varying a particular parameter across different values; then using that information to make an informed decision about which specific parameter value to use in actual trading. Using parameter values in actual trading that result from proper optimization should increase the likelihood of getting good results in actual trading in the future. A specific example will help. Consider the rules to the Original Turtle System, which I and others have made available free on the site: www.originalturtles.org. The "Unit Add in N" Parameter One of the rules is that positions should be put on piecemeal, each part was called a unit. The second part of a position was added at 1/2 ATR (Average True Range which was known as N to the Turtles) after the initial position, and subsequent parts each 1/2 ATR later. In the Turtle System, the 1/2 in the position addition rule is a parameter. Now, consider a test made using the VeriTrader? Demo of this parameter, known as "Unit Add in N". The following is a graph of the values of two measures of system performance, MAR and Sharpe Ratio, as the value of "Unit Add in N" varies from 0.0 ATR (meaning all the units are put on at once) and 1.0 ATR: Notice how the results for the 0.5 ATR value are the peak for this test. In fact, the results for 0.6 ATR are signification worse, the MAR ratio drops from over 4 to approximately 2.8. Unoptimized = Arbitrary Now, returning to our premise that optimization is beneficial, suppose we had not considered optimizing "Unit Add in N" and had started with a value of 1.0, a nice round number. We would have been leaving a lot of money on the table, and would have subjected our trading to much greater drawdowns than a 0.5 value would have provided. Not optimizing is simply leaving things to chance through ignorance. Having done the optimization, we now have a much greater understanding of the performance ramifications of the "Unit Add in N" parameter and how the results are sensitive to this parameter. We know that it is important not to wait too long, that if the markets move 0.5 ATR, we should act immediately and add another unit. We know that waiting even a little bit longer will likely result in decreased performance. Avoiding this optimization research because we were afraid of over optimizing or curve fitting would have deprived us of a good deal of useful knowledge, knowledge that could materially improve our trading results. The Flip Side: Decreased Predictive Accuracy Now consider a few more parameters (again we'll use results from VeriTrader Demo so the reader can experiment directly with these concepts): The "Stop in N" Parameter The Stop for the Turtle System was expressed in ATR from the entry point. The following is a graph of the MAR for various values of the "Stop in N" parameter: Notice how a value of 2.0 for the stop ATR shows the highest MAR at slightly more than 3.0. The "Max Directional Units" Parameter The Turtle System had a limit to the number of units which could be put on in a given direction, long or short, called in the VeriTrader Demo the "Max Directional Units". The following graph shows the MAR and Sharpe Ratio for various values of this parameter: The "Max Directional Units" value of 10 units is significantly better than any other value with the highest MAR and Sharpe Ratio. Notice the steep drop off between 10 units and 11 units. The "Entry Failsafe Breakout" Parameter One of the most important, but least understood rules of the Turtle System involved an early entry when the last trade had been a loser. In the event that the last breakout was a profitable breakout, there was a late breakout called the "Failsafe Breakout" in the Original Turtle System rules. This breakout was to be taken irrespective of the profitability of the last breakout (see the Original Turtle Rules for more details). This parameter exhibits a broader range of results values with the highest value corresponding to a "Failsafe Entry Breakout" of 65. The best value, it might be argued, is actually 60, since it sits in the center of the region of higher results even though it does not represent the highest value (65 has a better MAR, 55 has both a better MAR and Sharpe Ratio). The Basis of Predictive Value A historical test has predictive value to the extent that it shows performance which a trader is likely to encounter in the future. The more the future is like the past, the more future trading results will be similar to historical simulation results. The main problem with using historical testing as a means of system analysis is that the future will never be exactly like the past. To the extent that a system captures its profits from the effects of unchanging human behavioral characteristics that reflect themselves in the market, the past offers a reasonable approximation of the future, but never an exact one. The historical results of a test which is run using all optimized parameters represents a very specific set of trades, those trades that would have resulted had the system been traded with the very best parameters. The corresponding simulation results represent a best-case view of the past. One should expect to get these results in actual trading should the future correspond exactly to the past. It won't! Now consider the above parameter graphs, each of these graphs has a shape like the top of a mountain with a peak value. One might represent a given parameter with the following graph: Two Example Parameters If the value at point A represents a typical non-optimized parameter value, and the value at point B represents an optimized parameter, I argue that B represents a better parameter value to trade but one where the future actual trading results will likely be worse than that indicated by historical tests. Parameter A is a worse parameter to trade but one with better predictive value because if the system is traded at that value, future actual trading results are just as likely to be better than worse than those indicated by the historical tests using value A for the parameter. Why is this? To make this clearer let's assume that the future will vary such that it is likely to alter the graph slightly to the left or the right, we don't know which. The following graph shows A and B with a band of values to the left and right which represent the possible shifts due to the future being different than the past which we'll call Margins of Error: Two Parameters with Margins of Error In the case of value A, any shifts of the graph to right which would cause the value of A to move left on the graph will result in worse performance than point A, any shifts of the graph to the left will result in better performance. So A represents a decent predictor irrespective of how the future changes since it is just as likely to be under predicting the future as over predicting the future. The same is not the case with value B. In all cases, any shift at all, either to the left or the right, will result in worse performance. This means that a test run with a value of B is very likely to be over predicting the futures result. When this effect is compounded across many different parameters, the effect of a drift in the future will also be compounded meaning that with many optimized parameters it becomes more and more unlikely that the future will be as bright as the predictions of the testing using those optimized parameters. Important Note: This does not mean that we should rather use value A in our trading. Even in the event of a sizeable shift, the values around the B point are still higher than those around the A point. Now returning to the parameter "Unit Add in N": Note how the results steeply drop off to the right of the 0.5 ATR value. In the event of drift, a 0.5 ATR value is a somewhat risky choice for trading because of the risk that if the future was slightly different and the optimal value shifts lightly lower, there might be a significant drop off in performance of actual results corresponding with the drop shown here between 0.5 ATR and 0.6 ATR. The mitigating factor in this particular case is the fact that the 0.5 value is the original values given by Richard Dennis. It was optimal 20 years ago, and it has held up extremely well over many years. In fact, I can't recall a single test of the Turtle System of the hundreds or thousands that I have made over the years with many different markets, including stocks, where a value other than 0.5 had the best results. Factors that Affect Drift Several factors affect the drift (width of the Margin of Error) in historical tests results for parameter values over time, and hence the predictive value of optimized parameter tests: Number of Markets - Tests run with more markets will display less drift than those run with fewer markets. Tests optimized over the portfolio will have much less drift than those where the optimization has been done on a market specific basis. Amount of Data - Tests run over longer periods will have less drift than those run over shorter periods. Market Conditions of Test - Tests run over different types of markets will drift less than those run over specific markets, e.g. stock system tests run only over the last years of the bull market, compare with tests run over the last 20 years. For example, the VeriTrader Demo results are tested over a relatively small number of markets (15), over a fairly short time period (less than size years), for this reason the parameters are likely to drift significantly in the future, making it very unlikely that one would be able to achieve the results indicated using the optimized parameter values. Running the same tests and optimization process over more markets over a much longer period will generate results that are much better predictors of potential future results. Conclusion The Optimization Paradox has been the source of much confusion. It is also the source of much deception and scamming. Many unscrupulous system vendors have used the very high returns and incredible results made possible through optimization, especially over shorter periods of time using market-specific optimization. However, just because optimization can result in tests that overstate likely future results, this does not mean that optimization should not be done. Indeed, optimization is critical to the building of robust trading systems. In a future article, I will discuss how to improve the predictive value of optimized tests to compensate for the Optimization Paradox. --

※ 發信站: 批踢踢實業坊(ptt.cc) ◆ From: 218.167.220.57