I backtest those tidbits in trending and non-trending markets both bullish and bearish. You detrend the market, I backtest in all market conditions. I keep track of every trade I make, and what trading method I used. I basically keep score for each method and in what type of market it failed or was successful. If the tidbits have value that value will be reflected in my brokerage accounts.
Thanks. I gather you compute some performance statistic and then test to see if it is outside the luck band. When you do this for many such systems and consider eliminating some and retaining others, I gather you are using a test that takes into account the multiple comparisons involved. So really you are already doing what the book advocates. It's a case of been there done that from your standpoint. Got it. best David Aronson
It's amazing that the practitioners and pundits of esoteric Gann, Elliot, Fib levels, whatever, TA stuff don't realize that just applying a simple percent rule under/over low/high levels works as well, or better, as anything they espouse. So when the market dips/rises and their trade was profitable, it was because "the 5th wave coincided with the moon phase and Pluto is in Uranus, yada, yada..." But you could have done just as well with "Buy at limit 5% under last close..."
you gotta admit that elliot wave has STYLE though "it's the fifth wave of the 4th chewbacca formation with a dodecahedron sidecar..."
David: Although I have not read your book yet (arriving this week), Crabel's book is popular now because he was a trader who was able to parlay his research into real money, i.e., he runs a billion dollar hedge fund. The appeal of Crabel's book is that he enumerated hundreds of price patterns, combined them with daily open/close and range sequences, and calculated the P&L for all of these permutations. So, an inexperienced trader would be tempted to pick out the best-performing patterns (your "data mining bias") and probably lose his money. The brilliance of Crabel's book is his analysis of range contraction and expansion for volatility breakout systems. When I develop a system, I prefer to first validate a principle (e.g., mean reversion, volatility expansion, etc.) rather than program a rule. My market framework is much more important to me than a given system, as long as I have a choice of systems that work within that framework (e.g., high vs. low volatility). Still, your fundamental question has not been addressed in any of the trading literature. But suppose I find what I think is an edge. Do I really care whether or not it is a real edge if I have a system in place to monitor deterioration in the performance of the system that monitors the so-called edge? Maybe the "edge" stops working right away, maybe it lasts another year. Take the "open range breakout" for example. I recall a January several years ago where there was a trend day almost every day that month. I prefer to follow acrary's example of framing a day's price action (e.g., a trend day) and then mining the conditions that consistently identify trend days, using various time frames. I would tend to trust an edge that operates on a simple principle with less accuracy than an edge that has eight degrees of freedom (variables). My most consistent system is an upward-biased stock trading system that uses a moving average and a calculation based on standard deviation off a weekly low. Two fundamental principles: upward bias of the stock market and mean reversion. So, I hope your book addresses not only the fundamental question of a true edge, but the degree of reliability that can be expected based on the variables defined by the system, the breadth of the markets tested, and the testing methodology. The greater the variability, the greater the probability that the edge can be chalked up to chance. Comments?
You have raised some deep and relevant issues. The issue I deal with is quite fundamental....and it does take into account the amount of search that led to the discovery. Thus there is a greater bias if all 1, 2, 3.....up to 8 variable models are explored before picking a best. But again this can be taken into account as I explain in the book. However, the results of each system investigated needs to be kept to produce the required null distribution to test the best system found. That I think is the key. As the experiments show, data mining is a good thing and more is better, given sufficient sample size. Sample size is critical and the experiments show this clearly and White offers a proof which I cite. But given sufficient sample size, the more one searches, the greater the chance of finding a good system - but the bias generated by the search must be taken into account. I explain how to do this. hope this helps and hope you find the book to be of some value. David Aronson
Speaking of search, what is your opinion of the accuracy of minute charts vs. tick charts vs. volume bar charts vs. range bar charts?
Sorry I don't have any informed opinion on the question you raise about the relative merits of differnt types of charts. David Aronson
yes, exactly. i discovered that i could draw the "gann channel" almost at random producing the same results as one painstakenly calculated using all the esoteric material gann has to offer. regards, surf
Thanks, it gives me a basis to build on when I get your book. I believe that you can't get accurate objective reasoning from an environment that is constantly in a state of flux. If your data is viewed in such an environment then the results will be as varying as the environment itself. Sort of like trying to consistently thread a needle while hopping on a pogo stick.