LT/ST conflicting situations

[Andrei]

In other words, it is best to keep everything as it is right...?

Pascal, have you ever considered adding an outside element to the Robots, for example any of the existing market indicators, be they momentum, trend, breadth or any other, to further improve methodology? I am asking simply to know your opinion on whether you think Robots are fully self-sufficient, or you think theoretically they could benefit from such fusion?
(by the way, I am well aware that any change would require hell load of a backtesting)

Thanks.

[Pascal]

In other words, it is best to keep everything as it is right...?

Pascal, have you ever considered adding an outside element to the Robots, for example any of the existing market indicators, be they momentum, trend, breadth or any other, to further improve methodology? I am asking simply to know your opinion on whether you think Robots are fully self-sufficient, or you think theoretically they could benefit from such fusion?
(by the way, I am well aware that any change would require hell load of a backtesting)

Thanks.

Andrei,


When you develop a trading system, the issue at some point is to identify the weaknesses and then find solutions.
However, trying to include new indicators without knowing what you want to improve is not helpful.

Regarding to your first question: is it best to keep the system as is?
The first question is to find if there is an issue with the set of conflicting situations that I published.
What lies in the background of this set of data? There is something obvious that leads to the real question and thus the real improvement that I just tested in the past few hours.

I will write later about this issue.


Pascal

[EB]

Regarding to your first question: is it best to keep the system as is?
The first question is to find if there is an issue with the set of conflicting situations that I published.
What lies in the background of this set of data? There is something obvious that leads to the real question and thus the real improvement that I just tested in the past few hours.

Not sure if this is the "obvious" to which you refer, but there were a string of large winning trades that would have been missed at the beginning of the 2009 bull...just a few weeks after the QE1 announcement.

[Timothy Clontz]

The purpose of a model is not only to maximize reward but to optimise the reward to risk ratio. Higher losses could offset that ratio even with slightly improved reward.

[mingpan.lam]

Look like this filter is used to identify the short covering activity by setting a threshold for ATR. Is this filter ATR - average(ATR) ??

[Timothy Clontz]

Hi Timothy,

How do you measure your system's reward to risk ratio when you back test your system? I also focus on profit and max drawn down and I think measuring the reward to risk ratio may be a better choice.

average (Profit/drawdown)

Cheers,

Ellis

Ellis,

I measure average profit over maximum drawdown, rather than average to average. The problem with volatility is the human element -- there's only so much pain I can handle. :-)

Tim

[mingpan.lam]

Ellis,

I measure average profit over maximum drawdown, rather than average to average. The problem with volatility is the human element -- there's only so much pain I can handle. :-)

Tim

Thks Tim.

Very true. Something the backtest result is good but when the drawdown happens, you feel the pain !!

[Pascal]

Not sure if this is the "obvious" to which you refer, but there were a string of large winning trades that would have been missed at the beginning of the 2009 bull...just a few weeks after the QE1 announcement.

Attachment 11228

Exactly! Then, the second obvious issue is that the great majority of the LT/ST conflicts are on "buy" signals.
It means that the IWM robot had conflicting situations on buy signals, especially at the start of QE1 and also in the past weeks. The common points between these two periods is the high volatility. High volatility is linked to fear and to better shorting opportunities. Hence, the IWM Robot does not see that at specific points in the market life, high volatility was a consequence of forced shorts covering.

We can see in the list below - from which I erased the short signals - that the average ATR is 3.43% this is 1% higher than the average ATR for the past 5 years. This confirms that there is a volatility interpretation issue.



The question is now: how can we build a filter that would sort out the bad red trades and keep the good green trades. This is where I'd go to Andrei and ask if he has in his tool box some sort of indicator that can do such a job.

I found such a filter and used it to build another column.



I then sorted the results, seprating the positive from the negative of this new filter.
I believe that we now have an improvement over the existing IWM Robot, that takes care of Buy signals in high volatility environments.




Pascal

[Andrei]

We can see in the list below - from which I erased the short signals - that the average ATR is 3.43% this is 1% higher than the average ATR for the past 5 years. This confirms that there is a volatility interpretation issue.

Attachment 11232

The question is now: how can we build a filter that would sort out the bad red trades and keep the good green trades. This is where I'd go to Andrei and ask if he has in his tool box some sort of indicator that can do such a job.

I found such a filter and used it to build another column.

My answer would be, that honestly I have no idea! But I see that you found such a filter... So, what is that column to the right?

Great job, Pascal!

[senco]

Regarding measuring risk/reward of a system: Maximum drawdown is not a very good measure for system risk. Also it is not a good predictor for future drawdown. There is math logic behind the popular saying "the worst drawdown of a system is yet to come".
We look at backtests trying to get an estimate for a system's future behavior with an underlying assumption that the statistical characteristics will remain the same. But the maximum drawdown is not a good representation of the statistical distribution of trade results; it expresses just a single path of trade that occured up to this point.
A simple example: The following trade results could come from two systems with similar statistical characteristics:
A. +2 -3 +2 -1 +2 -2 +1
B. +2 -2 -3 -1 +2 +2 +1
In evaluating the potential risk of a system, measures that look at the statistics of trade results and not a single path would be more predictive. Downward Deviation, or Downward Deviation of rolling periods would be a much better measure for system's risk. Sortino, or even Sharpe Ratio would give a better prediction for future system risk adjusted return than (past profit) / (maximum drawdown).