I am not autotrader too, I test strategy just because I want to know some trading logic is Ok or not and how. Yes your idea is very clear; letting "these all come into play" is more closely to real result; So the path: 1, test strategy by back-test;2, if 1 is Ok, trade a little on market by real money; 3,Assess the strategy again! In the step 2 of this course, 'if 1 is Ok', how define this 'Ok', or when it is deserved testing this strategy by real money?
hmmm....okay. In this case, I think the other posters are right. In order to go ahead with real money, I think a person first needs substantial experience with the backtesting/autotrading platform, probably over a year. AND THEN, watch it on walk-foward sim. AND THEN, trade with real money, as small as possible. Autotrading real money with little autotrading experience could end very badly....proceed with caution......You really wanna know your strategy inside and out. And you really wanna know your autotrading platform inside and out. Along with the programming language. Lots of unforeseen logic errors often show up in live trading. Gotta be somewhat experienced to catch and debug in a timely manner. I will shut up now
Thank you for sharing your good idea which show a clear and pratical logic and method for confirm a strategy. Back-testing is just a feasible tool; good thing alse has bad aspect and vice verse; yes much logic errors/bugs would be covered by programming language; debug must be.
I have used MT4 about 2 years, but now I have make code to test strategy on Matlab about more than 1 year.
Best practice is to keep things relatively simple, if you can't memorize and very quickly explain the strategy, it's probably too complex. Every additional parameter is a potential curve fit. Walk forward testing can be useful but it implies you have parameters that you'll be optimizing. I look at market conditions as oscillating, that means once you optimize your strategy on the most recent data, the markets will change behavior and so you'll be lagging in performance vs. using static parameters. Realistically speaking, you'll probably be adjusting, readjusting and re-readjusting the strategy for years because as you follow its performance, you will notice new nuances and ways to improve. Most of my stuff is always work-in-progress and I don't think that's necessarily a bad thing as long as you don't make changes in a state of fear.
Think there's any utility to purposefully curve-fit strategies, that are routinely curve-fit and traded for short periods of time before updated?
What you said sound very nice; Best practice is to keep things relatively simple, but it needs to be adjusted, readjusted because of new nuances; does it will become more and more complex because of new nuances' readjusted? or 'readjusted' only means add some thing and reduce other things at the same time to keep simple or more simple. I don;t know whether you make program code to build/test your strategies; I heard some one said his strategy/logic is very simple, but built on by thousands of lines code or more. why? In addition, when you build a strategy, how do you confirm it is deserved your real money trading.
By adjusting, I don't always mean simple optimization of parameters but rather adding another conditional exit or running the strategy on a different instrument - something you didn't think of before and therefore could not have built into the optimization code. Granted, everything could be optimized but this means a lot of coding and quite a bit of processing power. Markets don't really change all that much so quickly. Big changes seem to come around once per decade. There is no clear answer here that I've found. You need to look at your overall strategy and decide if this is logical and adds to performance or is just a random fluke in a data series. There are many reasons why something is thousands of lines of code - for one, you need to handle various error scenarios and checks need to be in place that you're processing valid data. Secondly you might use code to for example find support/resistance levels or create custom indicators or process fundamental data, all this easily makes up thousands of lines of code. There's some gut feel involved when deciding if it's valid or a fluke. Some of the things I look at are the number of conditions, number of trades in relation to number of conditions, curve smoothness (Sharpe).
Thanks for everybody's advice; These days I am detecting my strategy and analysizing when it run well and when not; This strategy is a short term one, which means open order today and close it tomorrow; and I have tested it on about 2000 stocks, 80% of which is profitable. I find it runs well on stocks which have big trading volumn and price moves moderately, and not well on moving drasticly. Until now I still don't know why it is so exactly. Now I want to make some formula which can calculate oscillation's degree of the stocks' price; and then select all the stocks whose price move moderately for trading in the future. Shall I get some advice for its algorithm, or is there already formula like this?