Thats why it is an interesting fact. Many consistently profitable discretionary traders I have meet (including longer term ones) would have their analysis of their setups for the moment, but, almost always refer to "the tape being not right" as the reason to not take a trade, having a tighter stop for the current position, or, simply get out at once The more interesting thing is how much of this "tape being not right" rule plays in their success. If a trader taking 5 to 10 trades a day in a mkt, while the tape reading filter remove a lot of bad trades, or, as a minimum, making the trader having a tighter stop, faster scratch, then it becomes very significant. Lawrence
Interesting thread, bulat... Lawrence, please could you go into more depth on what you mean by "the tape not being right" (perhaps an example or two)... Thanks
The most common ones I've seen can be explained by the type of reversal happened at 3pm ET today in the indices, which is not a perfect test (made a sightly higher high for the day for YM and ES but not confirmed by the NQ). Being bullish, the more experienced traders usually will say something is not right with the tape and they are going to take, if not all, then at least most of the profit there at the day high (the one made in the morning) before the mkt even started to sink. That single reason alone made them override their usual practice and beliefs in letting the profit ride. That improved their bottomline drastically comparing to others who stick to their guns and ride til the end of the day. I am pretty amazed at these tape readers' ability to sense dangers ahead of time Lawrence
I think experienced tape reading (including tape "feel" beyond the intraday time frame) can be far more effective than simple S/R patterns, to the point where you can sell after resistance is broken and anticipate the far-side support breaking -- in other words, way ahead of the crowd. But it's nearly impossible to quantify or systemize market "rhythm" imo.
if it was not alan i would say that something like this: http://www.elitetrader.com/vb/showt...9&highlight=system+stopped+working#post249509 is the pure curve fit. many variables, all of them with only loose results by themselves added up. i believe alan is true but i can hardly understand why. i think his edge test is no guarantee against curve fit. someone mentioned out of sample tests. it all boils down to the quantity of trials. if you search out 1 billion set ups, you will find 100.000 good systems, if you put all of them on paper trading you will end with 10 that did well out of sample. hi, nassim, thanks for passing by. peace
I don't know how I missed this thread earlier. Good thread. Responding to the original post, your concern is a well known statistical flaw called the multiple hypothesis or multiple comparison problem. Using an uncorrected T test to evaluate the statistical significance of a large number of null hypotheses will inevitably produce misleading results. This is by far the most common statistical error in backtesting stock market returns. The multiple hypothesis problem exists even if no formal statistical test is performed at all; human intuition is subject to exactly the same flaw as the T test, overestimating the significance of a result that has been selected from a large pool of hypotheses. The multiple hypothesis problem can be addressed relatively easily with methods such as the Bonferroni correction: http://mathworld.wolfram.com/BonferroniCorrection.html The overwhelming majority of backtesting studies have deep and irreconcilable statistical flaws. If you design your backtest from the ground up to be statistically valid and testable, you will find it much harder to develop a good strategy but much more rewarding to trade that strategy. I can state with certainty from my own experience that data mining for trading patterns is possible and quite rewarding, as long as you pay proper attention to statistical validity. Martin
any experienced or semi-experienced systems developer should tell you that curve-fitting is the ultimate no-no in systems development, but what is curve fitting? this is a very ugly phrase because it's often related to connotations of fantasy and expected future failure, which is true. curve fitting is fitting tight pants that may only be suitable for one butt and nobody else. any dude with enough time on his hands should be able to create a false but pretty model for any single market using five or.... a few dozen variables that look magical in the past. that's not really hard to do. but when we're talking about basic methods of optimization or relaxed data mining, is that part of curve fitting, too? well, it could be if it's badly abused and used with shitty inputs, but i don't think all of it should be placed into the same categorization. there are more disciplined and conservative ways to use the tool. when you're working on finding optimal keys, once in a blue moon, what you have may not be a lame curve-fitted function but perhaps something that's just so unreal because it makes ridiculously good sense across multiple time frames in various markets. when you're working with a non-randomly generated model with a set of expected inputs, you should already have a general intuitive sense of what you're looking for anyway, so the process and results shouldn't always be so mysterious. yes, no one should ever automatically accept immediate outputs from this method, but one should probably not automatically reject all results neither because one might just miss out on something, even if it's just one or two single elements of the puzzle. if applied with proper judgement, it just might effectively help you pinpoint certain weak or strong areas of formations that you might really want to study. you are not aiming for exact precision but more on a practical territory on a map. your goal is really to investigate what works and what doesn't work, and what's realistic or not, so why limit yourself if you have that capability to do so?
You can reduce the data-mining problem by testing on different periods. E.g. test 1990-95, then take the best patterns from there and test them to 95-2000. Retain the ones that work and then apply them to 2000-2005. The ones that worked in all 3 periods are much more likely to be robust than if you just data-mined 1990-2005 at once.
Technically speaking, curve fitting is the goal. Curve fitting is the art of matching a theoretical model to the underlying behavior of a multidimensional data set without allowing random variations to influence the model parameters. By analogy, when backtesting we want to find a trading model that exploits a persistent market phenomenon rather than just the vagaries of historical stock data. The key distinction is between curve fitting and over-fitting where your model or strategy reflects market noise rather than a persistent, exploitable market phenomenon. People in this business loosly use the word "curve-fitting" as a synonym for "overfitting." It's not a big deal, but it does occasionally lead to confusion. Martin
uhm, yeah. excuse my semantics. curve-fitting just rings a rather cacophonous bell in the head, you know...