walter, I have always wondered about how things would work trading based on how the strategy had traded recently...... Interesting concept... It may be worth noting however that the strategy I now trade is a variation of this strategy....... It was adopted to fit the ER2 better...... the same period where this strategy for the ES sucked the ER2 version did better...... Part of the problem I think during that period of time was the volatility, its when all the volatility was sucked out of the market... but more volatility has returned making the ES profitable with it again.. the ER2 is a lot more volatile than the ES and I think that is why the strategy survives during that time with the ER2
I think walter has made a *very* interesting point about overfitting here, and I'd like to pick his brains on it a bit further. Let me give an example: about a year ago I was working with a moving average crossover system with about 4 parameters, as I recall. I had about 2 years of data (on Forex) and I applied a nonlinear optimization algorithm (downhill simplex with simulated annealing, taken from Numerical Recipes with a few mods) to optimize as much as possible. What happened was that the algorithm found places in the parameter space with stellar results, but it did it largely by reducing the number of trades to a very small number. So whilst it seems you're saying as long as the number of parameters is sufficiently small compared to the dataset, there is no need to do an out-of-sample validation, theoretically, do we also need to point out that the number of trades must be sufficiently large too? Have I understood you correctly?
the problem with statistical machine learning classification systems is that for every piece of sample data the hypothesis is either right or wrong. it is slightly different with trading since we generally accept 'not trading' as not being an incorrect answer. iv played around with the idea of modifying my optimization algorithms to weigh in 'not trading' as a negative so that it can find a healthy balance. with a well defined exit strategy. you can take every tick and say 'if i traded on this tick. would I have made profit?' and then use that to classify the tick as either a buy or sell. then optimize over the entire sample set keeping track of all correct and incorrect guesses, maybe scaling somehow by abs p/l. ie. a correct choice making 2$ is weighed twice as much as a correct choice making 1$. an incorrect choice making -$4 is weighed 4x as much as +1$, but negatively. i wish i could remember a theorem that addresses hypothesis space size. realistically if you are trying to classify 100 fruits as either apples or oranges, and you know it is roughly a 50/50 distribution. if you get a simple hypothesis that goes through a 100 fruits and only classifies 4 of them as apples. even though it might be correct 100% on those four guesses, it was useless for most of the rest of the fruits. and we can say this hypothesis is bad. since it misclassified ~50 - 4 apples. intuitively. there should definitely be a large number of correct responses (or responses in general) from a hypothesis before assuming it is a good representation. As stated in a previous post, I think since the market is consistently changing that the distribution of right / wrong hypothesis estimations has significant statistical value. A constant distribution or an increasing distribution leading up to real-time would show more promise than a decreasing distribution as shown in frost's 5yr graph. viewing draw downs as incorrect classifications. http://www.student.cs.uwaterloo.ca/~cs498/Lowerbounds.pdf "The first claim we prove shows that if the size m of the sample S is not large enough, then there will exist a âbadâ distribution whereby the real and sample error differ by some specified amount." the issue with most trading optimization is that the sample set gets pruned for every sample it decides it can not make a prediction for. in the end reducing sample size => reducing statistical validity. note we cannot use the equations in the link for markets since we cannot prove they are not PAC learnable <=> infinite vc-dimension. one can always hope (gamble) though.
You may want to study this system...and may even want to contact the owner, Mike Barna....he's been around a while. His Big Blue system despite it's sophistication gets trapped during flat, low volatility periods as well. What can be done when this occurs ? That's a good question...."stop trading the system" could be the answer. http://www.tradingsystemlab.com/files/THE BIG BLUE2 TRADING SYSTEM.pdf
Lost -$344 today... I am getting tempted to pull the system and run it on sim mode until it shows again that its becomming profitable..... or at leat until the first of january when I think the market will start to move well again....
Not surprising. Your system is not programmed to adapt to lower volatility conditions or flat, sideways markets. Yeah, I know...everyone says "make it simple, simpler-the-better, etc, blah, blah, blah....". Well that applies to longer term EOD systems, but for tackling shorter time-frames, more sophistication is needed. What the heck, why would all of these hedge funds be only looking for PhD candidates for building their high-frequency strats ? Adaptivity is the key. How to do it ? Wow, that's a whole new can of worms !
Am I reading this pdf right? It states on pg. 9: "the hypothetical results do not include slippage and commissions". I wouldn't bother with the rest of the document.
from a wise system trader: "A more common expression of this phenomenon is how some individuals can prove a strategy through comprehensive backtesting, only to dismiss it as a failure after trading it in realtime for as little as 10 trades. It appears they just don't have the stomach for probability. Their ability to maintain their 'Mathematical' objectivity is lost because their hip pocket nerves eventually override all other thinking by transmitting ever increasing pain signals to the brain. It is a well accepted fact amongst professional traders that a losing streak of 8 consecutive losses is not at all uncommon. This is why the ability to sustain losses is so important, not only in terms of your trading capital but also with respect to your 'Psychological' capital. To preserve your trading capital, I recommend sticking to the 2% risk rule. To preserve your psychological capital I recommend carrying a coin at all times; you never know when you might have to play a game of coin toss and prove to yourself that a string of 8 consecutive losses is very possible"
I've been reading this book "Introduction to High Frequency Finance", and I've been thinking about why hedge funds want PhDs. Really, it is all "math elitism." I'm convinced -- no, in fact, I'm certain -- none of these high frequency systems require more than a competent electrical engineering student with a solid signal processing background with a few extra probability and statistics classes. If you look at the way the book is written and who it's trying to target, it all starts to make sense. Rather than state the obvious to anyone who's done any signal processing, they take great pains to talk about convolution integrals in the most convoluted way possible and never state anything about the general idea. At times, I think the author is trying to show off math skills rather than the core ideas. Wall St. is the same way. So in short, there is greater sophistication but it certainly isn't PhD level work. The PhD thing is just Wall St.'s way of doing everything possible to filter out people who don't have a grasp of the basic math concepts.
read either pdfs iv posted on recent pages and say that again. and remember those are both from 'introductory' machine learning courses. basically if you want to get a high paying job researching/programming ai on wall street. you need to know at least that + some. that being said. iv seen job postings for phds in stat arb that were starting salary of +8,000,000$. it is definately not a way to keep 'non mathies' out.