A good strategy typically works in most liquid instruments or most illiquid instruments and in some cases, both.
Should Strategy Work Across All Instruments
- Thread starter Dhalsim
- Start date
Thinking further about the above, I suppose some edges seek to exploit some specific aspect of trader behavior/psychology (rather than some specific aspect of market structure).
In these cases, to the extent that the same traders are trading various instruments (and to the extent also that these traders continue to behave the same way when doing so), such an edge ought to work on all those instruments.
So, in some cases it would be correct to expect the egde to work in many instruments. In other cases, it might not be. It depends.
Curve-fitting IS a bad thing because it does what you described, it maximizes hypothetical profit. There is no link between curve-fitting on past data and higher future expected profit that I know of, unfortunately, or else the world would be filled with rich trader.
The issue is whether curve-fitting can be avoided and how. I think this is a good article on this subject that partly agrees with you in that curve-fitting may not always be the problem. I think IT IS always the problem.
IMO, any trading strategy is an exercise in curve-fitting.
The example provided in the type-III section of the blog oversees a key aspect:
to get to the entry rule "if Close[0] > Close[2]", one has to datamine and find that this is the best entry signal over many candidates in the form of "if Close[0] > Close" where i IS an optimization parameter (not to mention a gazillion other variants using High/Low/Median/ etc).
Even in the case where absolutely no datamining is involved, and someone just decide to test a strategy based on a perceived market dynamic, that strategy will be tweaked & ultimately accepted or rejected based on past performance (even forward testing, at the time of making a decision to accept / reject a strat, amounts to looking at past performance).
The question should be: where is the (fine) line between proper curve-fitting and self-destructive curve-fitting (aka over-fitting). I believe a great deal of it resides in the sample size considered to begin with. Get 3000+ samples of a given pattern, and it becomes easier so see if it has an edge better than random. Another aspect I like to analyze for each candidate pattern is its SQN (avg / stdev), and how that metric varies year over year. A high & stable SQN for a 500 samples pattern can be just as good as a lower & less stable SQN for a 3000 samples pattern.
The example provided in the type-III section of the blog oversees a key aspect:
to get to the entry rule "if Close[0] > Close[2]", one has to datamine and find that this is the best entry signal over many candidates in the form of "if Close[0] > Close" where i IS an optimization parameter (not to mention a gazillion other variants using High/Low/Median/ etc).
Even in the case where absolutely no datamining is involved, and someone just decide to test a strategy based on a perceived market dynamic, that strategy will be tweaked & ultimately accepted or rejected based on past performance (even forward testing, at the time of making a decision to accept / reject a strat, amounts to looking at past performance).
The question should be: where is the (fine) line between proper curve-fitting and self-destructive curve-fitting (aka over-fitting). I believe a great deal of it resides in the sample size considered to begin with. Get 3000+ samples of a given pattern, and it becomes easier so see if it has an edge better than random. Another aspect I like to analyze for each candidate pattern is its SQN (avg / stdev), and how that metric varies year over year. A high & stable SQN for a 500 samples pattern can be just as good as a lower & less stable SQN for a 3000 samples pattern.
One question, many different answers... What is curve fitting? Every methodology is a curve fitted system to you and what you believe about the market you are trading...
Practical: No, not one strategy should work on all instruments, all the time. You need a top down approach to know when to use a strategy. Volatility and direction can be used to determine wich strategy to use. An instrument can behave quit different in a low volatility, bull market then when it's in a high volatility, bear market. Do not expect your strategy to perform the same in both condition, even with one instrument...
Practical: No, not one strategy should work on all instruments, all the time. You need a top down approach to know when to use a strategy. Volatility and direction can be used to determine wich strategy to use. An instrument can behave quit different in a low volatility, bull market then when it's in a high volatility, bear market. Do not expect your strategy to perform the same in both condition, even with one instrument...
I think this is confusing question. You minimize the possibility of curve-fitting by determining that the strategy works over many many instruments. If it does not it may be curve-fitted or it may be something else. But if its works it is not curve-fitted. You do not care to prove it is curve-fitted, you care to prove it is not or it is minimally such.
I agree. I also think the author agrees that is why he calls it a type-III curve-fit.
By the way I think his program price action lab has a lot of potential if properly utilized because dealing only with price patterns minimizes data-mining bias due to the similarity of rules. That is also what Aronson claims in his book, i.e. that data-mining bias increases as you add more dissimilar indicators and decreases otherwise.