Wipe off your screen ALL the way down to its surface. Just mark the end of the running with a China pencil. 28 that's really getting squirelly
backtesting is stupid
- Thread starter jonp
- Start date
Wipe off your screen ALL the way down to its surface. Just mark the end of the running with a China pencil. 28 that's really getting squirelly
Was wondering whether someone already tried to backtest a random walk. 
Seriously though, one could create the equivalent of many years of "quasi random walk" time series (say one introduces brief periods of trending data, ie non fully random, just like markets do every now and then - but does so in a random fashion, needless to say
).
Then take all that data thru their preferred backtesting/trading system dev software tools, see if they can find an ""edge"" over that random time series...
I wouldn't be surprised if one can find "edges" within random data.
Then of course comes the slight issue that there is close to zero chance that the said ""edge"" would yield any similar results in the future... :-[
The next question then would be: based on the (tbp) fact that edges can be found even within random walk type time series, should we all stop considering backtesting a valid strategy?
Seriously though, one could create the equivalent of many years of "quasi random walk" time series (say one introduces brief periods of trending data, ie non fully random, just like markets do every now and then - but does so in a random fashion, needless to say
).Then take all that data thru their preferred backtesting/trading system dev software tools, see if they can find an ""edge"" over that random time series...
I wouldn't be surprised if one can find "edges" within random data.
Then of course comes the slight issue that there is close to zero chance that the said ""edge"" would yield any similar results in the future... :-[
The next question then would be: based on the (tbp) fact that edges can be found even within random walk type time series, should we all stop considering backtesting a valid strategy?
I do not think that relatively I use a longer history, as I use 15 min bars and need about 1 year of data or at least a period which would include uptrend, downtrend, ranging, all spectrum of volatilities and volumes. And I see nothing wrong with your weekly re-optimization if it is not over-optimization and improves your results. In general I am always looking for a system where if you even enter at least favorable parameters you'd still be positive in long run.
Probably 1-min of 6-months history should be sufficient particularly if you feel that your algorithm is not a random play of the data but based on your observation of the market flow.
I am sure if you create a simulator which generates truly "random walk" data, then you will be much successful with the martingale approach.
Quote from cfd_trader:
Was wondering whether someone already tried to backtest a random walk.
Seriously though, one could create the equivalent of many years of "quasi random walk" time series (say one introduces brief periods of trending data, ie non fully random, just like markets do every now and then - but does so in a random fashion, needless to say
Then take all that data thru their preferred backtesting/trading system dev software tools, see if they can find an ""edge"" over that random time series...
I wouldn't be surprised if one can find "edges" within random data.
Then of course comes the slight issue that there is close to zero chance that the said ""edge"" would yield any similar results in the future... :-[
The next question then would be: based on the (tbp) fact that edges can be found even within random walk type time series, should we all stop considering backtesting a valid strategy?
Well, duh! Your 5 years of 15 minute is not that much more than my half-year of one minute. Thanks for your comments on my time period. But I am having trouble wrapping my old head around "if you even enter at least favorable parameters you'd still be positive in long run." In what I do I try to ruthlessly drive out "least favorable parameters" to reduce the number of losers relative to winners, even if that means a lower net profit. For example I have one system where I could net 20% more but I would gag at increasing losers by a factor of 30% (roughly four per win rather than three).
well as your indicators told you that earnings would not be good, i'm sure you made a ton of money shorting it.
on a serious note this stings for a few reasons. I had a profit which I did not take this afternoon and was one click away from selling above 130. also because I posted it here. also because my analysis was just wrong, i figured intel just beat, amd did well, oracle did well so it was ibm's time to make 5-6% after hours. anyway now I am stuck with this shit
Maybe the way you backtest is stupid. I suggets you read a few good books on the subject. Here is a free but good article