window / number of trades to evaluate performance of a system?

[MisterH]

I'm looking for ideas on how to decide what the appropriate size would be of the window to evaluate the dynamic performance of a system on.

- One idea would simply be to take the last 20 trades and run your statistics on that to decide whether the system still works fine in the current market.

- Another idea would be to base it on the best/worst historical sequence of trades: take the win% of the past 20 trades,

the maximum number of consecutive winning trades over all trades,

and your window length would be 1.5 * (max conseq winners / win%(20))

(https://championship.mql5.com/2012/en/2008)

- However, there are views that consecutive winners/losers don't mean anything because trading a profitable system is like flipping a coin that is

weighted on 1 side. You're inevitably going to have winning and losing streaks, but they are random.

So pls some other ideas

[blogzr3]

Depends on too many factors, the most important one of which is your actual strategy.

- a sample size of 20 is meaningless, statistically speaking (the article did say it's a randomly chosen number)

- simple ratio of winning/losing trades are meaningless, if you have 19 small wins and 1 big loss to wipe out the 19 wins.

Winning streaks and losing streaks are most definitely NOT random. They can be quantified, in terms of statistical probabilities. The trick is knowing how.

[MisterH]

blogzr3:

Depends on too many factors, the most important one of which is your actual strategy.

- a sample size of 20 is meaningless, statistically speaking (the article did say it's a randomly chosen number)

- simple ratio of winning/losing trades are meaningless, if you have 19 small wins and 1 big loss to wipe out the 19 wins.

Winning streaks and losing streaks are most definitely NOT random. They can be quantified, in terms of statistical probabilities. The trick is knowing how.

-if a sample size of 20 is meaningless, then what is in your opinion meaningfull? 35 observations per degree of freedom? (95th percentile of a chi-squared statistic).

Or how else would you calculate minimally required sample size (length of the window)?

Running simulations and checking if the 95% confidence interval of the mean has indead 95% coverage?

Please mind that the article only used the last 20 trades to calculate the window length on, not to calculate any performance statistics like win%, PF, ...

-I did not say that winning & losing streaks are random. I said some people say that consecutive winners/losers don't mean anything.

Pls some actual suggestions