World Finance 100
LiteForex community on Facebook LiteForex community on Twitter LiteForex community on Google+ LiteForex community on Youtube RSS Feed LiteForex community on MQL5

Settings (adjustment)

Analysing the results of your algorithm testing. Adjustment to the received results

In accordance to analysis results of the test a trader can make a decision about the changes of those or these features with the goal of optimization or maximization those or these features of trade strategy. In particular, you can try to maximize profit or find optimal correlation of profit and risk at the current level of risk.

The most important nuance in the question of trade strategy - is segmentation of all available massive of work of historic data into two general intervals: test and valid one. The size of valid interval should give representational statistical sampling. The size of test interval can be larger or the same (it is desirable for this size to be twice larger).

Picture 19. Test and valid interval

The deal is that any given historic period can choose those settings of the strategy that will make it more profitable. The strategy in process of such setting is "matched" to the given data. The effect of so called "resetting" is when the strategy with definite settings gives good profit at the test interval, but at the valid interval ("unknown" for the trade work) the profit goes down.

On the basis of Turtle strategy let us see what settings can we change in process of optimization:

1. "Sensitivity" of the system to appearance of new trends increases, of the entering follows the break of less important levels (for example, instead of 20-hours maximum or minimum you can use 10-hours one). This will lead to the situation when the system will enter the market more frequently, and it means, that it will make more transactions. n the one hand, it is a plus. However on the other hand this more "sensitive" system will react to the bigger quantity of false breaks, and thus, the quantity of unprofitable transactions with those of them that follows one another, will grow. If the "sensitivity" of signals become smaller, the system will catch more significant trends. However in this case the system will enter the market significantly rarely and most of the time, the funds will not work, and will just stand idle.

2. In the same way you can change the "sensitivity" of the system to the turns, decreasing the degree of importance of extremums that give signals to quit.

3. Changing the capital share that is given for each certain transaction, you can increase the efficiency of money usage (i.e. the capital percent that makes part of investments will increase). However you should take into consideration that this increase of some long series of losses can bring strategy to deposit loss.

In case if the change of those or these settings bring to the successful results of the work of system at the test interval, the system should be tested at the valid interval to make sure that there was no "resetting" to the concrete historic period. The system, the settings of which are chosen optimally will give approximately the same results during testing both at test and valid intervals.

Apart from this the similar settings should be done (choosing test and valid intervals from the whole available historic period) for several different currency pairs. Thus the system will work normally with forecasted results for the chosen currency pairs (In our example it is the group of "margins"). 

It is strongly unrecommended to fudge settings of the system to the valid interval. In cases when you see that the system changes its behavior sharply at the valid interval, you should make a conclusion about probable "resetting" of the system to test interval and make a step back. 

Among other mistakes it has to be noted the situation when very good results of system work are obtained only at concrete test interval. The market dynamics at this interval can be very suitable for your strategy. For example, for Turtle strategy it is a very long trend with insignificant drawdowns. Even if test period is very long, but within this period the market behavior was characterized as monotonous (for example, trend only or flat only), you should try another interval and estimate the given difference. You should not try to chose the optimal settings for each definite period: in future the market will make such a position to which the system will not be ready.