There must not necessarily be a reason, such as a sudden market change. There is also the possibility that the performance gap is just a random fluctuation of the distribution of your trade results. Finding out if it's a market change or a fluctuation is relatively simple, by generating a set of Monte Carlo resamples of your equity curve, and checking the resulting curves for similar performance gaps. Here's a link to a paper: http://www.algodude.com/monte-carlo-resampling-of-equity-curves-using-n-bar-segments/ However, from what you describe, I suspect that the gap is indeed caused by a market change and will likely continue. You're trading price action, which means your results depend on very temporary and volatile market properties. One could also say that price action is just trading noise. This does not mean that your strategy can not be profitable, but it might be short lived, especially since you said that you manually tweaked the strategy parameters a lot. For making such a strategy stable, you could formulate an optimization algorithm for those parameters, and then do a WFO optimization over a lot of assets and time periods. This could give you at least some good information if you should continue to develop and trade it, or give it up.
As for me personally, I did not imply that systems trading does not work. I specifically meant that the OP cannot and will not code what he sees into long-term sustained, profitable success. That is solely do the the fact that 1-2-3 type patterns of any descript have so many variables to unfold, it is impossible to define by code. For example: many of these type setups finally work on the second or third thrust. How do you instruct the computer to manage trades that chop once or twice before finally taking off for substantial gains in favor the third time? Those are some differences between human observation and hard coding into computer language. I have plenty of experience writing and hiring experts to write similar language to have a valid opinion. Whether anyone cares to heed or ignore is strictly up to them
pretty much what it is... classic traps = springs All of those off-shelf systems are long-term failures in real-time trading, too. Not the fault of setups or logic... purely limitation as to what computers can be instructed by someone with limited resources trying to create simple bots from complex equations
I agree there are styles (and aspects of trading) where the (properly trained/experienced) human mind has major advantages over the computer. I think the optimal solution is to meet somewhere in the middle (essentially the computer/stats/systems/mm/risk management/etc are all "decision support tools"). Both extremes (super-systematic or pure discretionary) can work though in the proper hands, and it becomes a matter of personal style/preference for most. I'm personally much more mechanical than what I envision a hybrid trader would optimally be, and thus I suffer the consequences of sub-optimal performance compared to such a dream trader, but I decided a long time ago to focus on robustness and survivability vs. get-rich-quick or having to camp the screen constantly. This young trader/OP might benefit from looking for edges/systems that are perhaps not holy-grail in appearance but seem to be robust, and keep adding on (preferably as little correlated as possible) edges/systems to build essentially one big portfolio of systems that are much better in sum than they are apart. This is how I make up for the faults of my fairly systematic approach (the discretion is more in how systems get turned on/off and what level of risk to take, not in individual trading decisions themselves). Focus on finding a wide range of acceptable answers to the problem and use them all rather than trying to find one master key to unlock the answer...
Yes, there are: fundamental analysis. A computer won't care about the latest news, unless it's in a novel by Robert Harris. Technical trading is primitive. No matter how much intuition or gut feeling you use, a computer can not only do it, it can do it far better.
jcl, you are spewing a crock of newbie EA bullshit. crank out 1000 trades with your machine code and we will reconvene in 6 months. a real account machine traded p and l blotter is required. consider yourself corrected. s
Do I understand this right: you're going to admit superiority of machine trading when I show you my blotter?
I wanted to thank you all for your constructive feedback ... at this time I have busied myself reviewing the differences between timeframes (V180 / V200 / V220 ... I added V100), which are essentially coming near the thresholds for pivots identification ... varying the timeframe sometimes changes the number of bars to the right or left of a pivot, making it appear or disappear "randomly" (I witnessed one pivot, present on the V200, but missing on both V180 & V220 for that very reason). There isn't much I can do about it, I have started migrating to V100 which will provide a tiny bit more resolution. All settings adjusted to have as similar thresholds to the V200 as possible, avg/trade varies from 76% on the V180 to 94% to V100, through 85% for the V220 (those % are in reference to the V200 avg/trade). Again I agree it is highly probable there is some over-optimization, as the original timeframe used for the system development has clearly the best performance. With that said, there is also likely an edge in the system, not as big as the V200 performance would indicate, but I am thinking 75% of that ... I will redo the drawdown analysis using 75% & 50% of the original performance. I used the boostrap util from the link TD80 provided (thanks!), the p-value returned using the entire backtesting distribution (normalized to %) is 0 (zero), where for the entire drawdown period it is 0.39 ... not sure what to do with that. But the Aronson book just arrived today (I had ordered it a couple weeks ago), may be after reading it I'll have better ideas on this topic. I also realize I need to change trading platform for one better suited to system development & backtesting (currently using Ensign) ... but given the amount of work involved, my current line of thinking is to finish this system on Ensign, then take some time to select a new platform & get up to speed on it.
The p-value result shouldn't be 0 but it might be a really small number (and thus be showing 0 in excel or somesuch if not properly configured). Also keep in mind that the first page on that topic is talking about testing equity curves using a single rule. If you have a complete multi-rule system then you should refer to the follow-up article: http://www.automated-trading-system...ta-mining-bias-code-and-using-geometric-mean/