過度の最適化とターゲット設定戦略の長期的有効性回避(SQ)
17 replies
Everything we do in SQ is the epitome of curve fitting, data mining, and over optimizing. We are data mining a strategy by randomly generating the best strategies from just 1 asset’s data. We are curve fitting it based on 1 set length. Its already optimized from just those 2 previous builds, then we optimize it even more! I try to explain how to limit it has much as possible below. Professional systems are built to trade across a set of markets (futures/equities/fx) all using the same exact parameters with similar results in each market and last for decades. Thats robustness. We are not making “robust” strategies so it is of the utmost importance to be very careful.
This is the next video I’ve made. For education purposes only.
“This video is about avoiding over optimization, lightly covering ways to increase robustness, avoid certain common pitfalls, extend strategy life and increase probability of survival using StrategyQuant.”
NEW HD VERSION LINK:
https://www.youtube.com/watch?v=A5yNfnF2mZM
I cover data length, spread size, slippage, step sizing in optimization to avoid curve fitting, knowing when to reoptimize a strategy without using walk forward (and more accurately) and much more.
楽しんでください。.
Comment here. Share your ways too. This is just mine. Part of doing this is to start a discussion and have other SQ users input their logic of doing things.
Not sure on what video I will do next, probably money management or robustness testing and reading results. Comment any requests as well.
Someone commented in my in video to make a public “MYFXbook” account to verify my EAs. I will make a new account specifically for public use. (my other accounts are for my own private use only). The public account will just be a simple tool to prove that htey are working. Probably put like 1500-2k$ in it and will add all the EAs I make in SQ and EA wizard to it. The same ones i’ve had running live for 1 year, and any new ones I make.
0
こんにちは、Threshold,
Thanks for the video. 🙂
Just watching it now. One small comment, the quality of 480p is too low to read anything properly on my HD screen. If I make the video small, it’s too small to read and if I make it fullscreen, it’s too blurred to read.
So I’m following what you say rather than what’s on the screen. I think 720p would probably be the min in order to read the screen.
乾杯,
マイク
0
Ok I think its too long for youtube. I will need to split it into 2 parts probably. Give me a day.
You are right, the other 2 videos have HD option this does not. Definitely will fix.
0
Hello Scott,
wouldn’t you be interested in publishing these videos also on SQ website in some section?
I think they’ll find many more watchers there.
マーク
StrategyQuantアーキテクト
0
Actually I am building only systems that work on all currencies (7 currently as SQ doens´t allow more right now). And they are working going foward. So I doubt I am curve fitting as you say in your initial post;)
0
Additionally each system found this way has to pass all kind of robustness tests too, just to make sure;) Oh and these 7 pairs are all on 14 years of data. And yes, you need a powerful machine and heaps of RAM for that. Rented here for that: https://contabo.com/?show=servers (the X and the Dual Xeon are good options), doesn´t get much cheaper than that for this hardware and they don´t mind if you run on 100% 24 hours too:)
0
Wow nice. Yes definitely takes some computing power.
I have a similar robust setup of 2 strategies made in EA wizard (1 trend follower 1 mean reversion) that run on all the majors+a few crosses+ metals & oil and are able to run on several other commodities as well with all the same parameters. The logic and pattern set up tends to make more sense too when human designed. They trade d1 bars so, minimum testing was 15 years when I made them but even on H1 bars I like to use ~ 14 years for random gen.
0
HD just added.
https://www.youtube.com/watch?v=A5yNfnF2mZM
0
Wow nice. Yes definitely takes some computing power.
I have a similar robust setup of 2 strategies made in EA wizard (1 trend follower 1 mean reversion) that run on all the majors+a few crosses+ metals & oil and are able to run on several other commodities as well with all the same parameters. The logic and pattern set up tends to make more sense too when human designed. They trade d1 bars so, minimum testing was 15 years when I made them but even on H1 bars I like to use ~ 14 years for random gen.
Totally agreed and good approach. Almost the same here, just that mine come from SQ and are neither trend followers nor mean reversers. Still they are totally simple, 2 indicators, rest comes down to good trade management, they are not about the entries, but cutting losers short and let winners run. And logically they make sense too, just never thought of that indicator combination. That´s what SQ is really useful for!
Keep up the good work with your tutorials, hope a lot more people can become profitable through them as many seem to struggle here as they go on it with the wrong approach.
0
Hi geektrader, which 7 pairs do you build for?
I’m assuming…
EURUSD
ポンド米ドル
米ドル/円
豪ドル/米ドル
NZDUSD
Then what?
0
No, uncorrelated ones (previously checking this with an extra software), so chosing the ones which correlate as few as possible so that the resulting system is as stable as possible and can handle everything. Those are currently (all have 14 years of data):
EURUSD
USD/CAD
EURJPY
GBPCHF
豪ドル/米ドル
NZDCAD
NZDJPY
0
Wow! And you’ve made strategies that are profitable on all of them??? I’m very impressed!
0
hey been long time no see. your profitable trader now huh.
long time no hear your voice lol.
nice videos btw.
0
ありがとう。.
Too bad you sold your keys!
0