patterns in error terms

[wonghang]

Hi all,

Actually, I have a strategy running for a long time.
I observed some kind of patterns of my error terms, I found that they appeared to have some kind of patterns.
I don't understand why. I am not good at statistic actually.
I performed dwtest() in R and could reject the null-hypothesis that the error terms have zero autocorrelation.
In fact they have +ve correlation. i.e. win trade is more likely to be followed by a win trade, loss trade is more likely to be followed by loss trade.
Do anyone get similar result when you trade your strategy everyday and have some insight on these error terms?
I want to know the reason why I will get +ve autocorrelation error terms.

 

[jcl]

Yes, we have similar results. In most medium and long term strategies, the sequences of winning and losing trades have no Gaussian distribution, but are autocorrelated.

The reason is simple - markets change within a longer time period than the average time between two trades. Those changes cause winners and losers to cluster, and generate a pattern in your error terms. You can take advantage of that in your strategy when you automatically reduce the trade volume slightly after losing, and increase it slightly after winning. Often this improves the overall profit.

 

[ssrrkk]

Quote from wonghang:

Hi all,

Actually, I have a strategy running for a long time.
I observed some kind of patterns of my error terms, I found that they appeared to have some kind of patterns.
I don't understand why. I am not good at statistic actually.
I performed dwtest() in R and could reject the null-hypothesis that the error terms have zero autocorrelation.
In fact they have +ve correlation. i.e. win trade is more likely to be followed by a win trade, loss trade is more likely to be followed by loss trade.
Do anyone get similar result when you trade your strategy everyday and have some insight on these error terms?
I want to know the reason why I will get +ve autocorrelation error terms.

Actually, when you have residual correlations in your error terms, then I think it means that your model is not fully capturing the correlations in your time series. Have you looked at the ACF or PACF? What time series are you looking at?