Topic: Explore the minefields of the optimization of the trading system (Read 146 times)

article originally from fmz.com         https://blog.mathquant.com/2018/12/12/explore-the-minefields-of-the-optimization-of-the-trading-system.html


First, optimization research restrictions

At present, some data providers still have not been able to provide a way to optimize transactional measures with multiple varieties. In fact, we all know that the optimal combination of parameters for each variety is not necessarily the same as the best combination of parameters in the trading strategy. And even data providers offer optimization and testing methods for a wide variety of parameter combinations. However, since the trader may also adopt some system combinations with negative correlation or low correlation, the optimal combination of parameters thus found is not necessarily the global optimal parameter combination.

Second, to avoid defects

We have already told you in the previous content that the most prone to system development is the curve fitting. In general, curve fitting is divided into two types, namely: parameter curve fitting and data curve fitting. Simply put, the data curve fitting is that when the system developer studies the filtering loss transaction most, it will consciously remove some data. This method is called data curve fitting.

In order to avoid such problems in the development of programmatic trading systems, traders can adhere to objective historical data standards while testing the system. This is the same as the problem we mentioned earlier. The market type in historical data must include the following: bear market, bull market, trend market and mean return market.

Of course, if we do not include these market patterns in the historical data, we must add more historical data. So what if we can’t find more historical data for a certain trading target? This is where we can find some other varieties for reference. However, we must note that the varieties we use for reference not only must contain data of each type, but also have a strong correlation with the original variety.

Curve fitting is where the system developer adjusts the parameter values ​​to match the data of the test. Let’s use a simple file to drive an example of an average crossover system. There is only one other parameters in this system, and we are very confident about the future performance and historical performance of this system.

However, if the developer is not satisfied with this relatively low win rate, then the developer may try to combine a series of parameter combinations until a relatively high combination of parameters is found without affecting other aspects. Then there may be two things happening at this time. The first is the reduction of the signal in the system; the second is the increase in the proportion of the profit signal in the residual signal. Then the developer may be able to make the system more successful by adding another parameter.

The end result is that the extra parameters added by the developer undermine the originally robust and successful programmatic trading system. Although the system is perfect for past data performance, the future data is really embarrassing. Therefore, traders must remember that if there are more parameters in the system, the possibility of parameter curve fitting will be greater. The more the parameter combination matches the historical data, the worse their adaptability to future random data will be.


article originally from fmz.com         https://blog.mathquant.com/2018/12/12/explore-the-minefields-of-the-optimization-of-the-trading-system.html