Is a trading strategy ever complete without using machine learning?
Recently I have seen a lot of traders training more and more ML models everyday, but I am also wondering if this has always been the case for creating trading strategies. I know that machine learning models are good for regime detection, and identifying anomalies (among other things).
So, can these improvements be made using pure rule-based algorithmic approach or is machine learning something that cannot be avoided when creating a truly robust strategy. What are your thoughts based on what you have seen or heard?
Personally, I avoid implementing complex models into trading. Generally speaking, I follow the KISS principle and Murphy's Law which suit me just fine.
I have however, considered using adaptive indicators─one step below machine learning, if you will. Anyway, I've never actually traded one.
Personally, I avoid implementing complex models into trading. Generally speaking, I follow the KISS principle and Murphy's Law which suit me just fine.
I have however, considered using adaptive indicators─one step below machine learning, if you will. Anyway, I've never actually traded one.
Great insight. I also think that when a model, or a trading strategy in general, passes a certain level of complexity, it recedes back to being unprofitable.
I love adaptive indicators too, but if you never traded one before, how are you sure its going to be effective? Backtest results perhaps?
Great insight. I also think that when a model, or a trading strategy in general, passes a certain level of complexity, it recedes back to being unprofitable.
I love adaptive indicators too, but if you never traded one before, how are you sure its going to be effective? Backtest results perhaps?
My initial stage of determining whether or not an indicator is valuable is running it through the Tester to merely visualize it on my desired symbol/chart. As they say, "a picture is worth a thousand words," and a moving chart is more like a moving picture. That works for almost every indicator, e.g., adaptive, repainting, etc. I simply have not seen a readily apparent additional edge in my testing adaptive indicators.
At this point, I'm a bit skeptical of replacing "standard" indicators, that already have an edge over the course a 5 year+ test, with adaptive indicators. I've found that constantly changing the parameters/logic of an indicator in real-time based on very recent live data is somewhat of a crapshoot.
I don't want to confuse adaptive indicators with repainting indicators. Some adaptive indicators can repaint history, but not all adaptive indicators repaint history─the latters simply adapt going forward.I don't think ML is unavoidable at all. Plenty of people trade rule-based systems that have held up for decades, and most of the ML I see posted here is a linear regression with extra steps.
Where I think it actually helps is when you have a lot of inputs and no clean way to combine them by hand. Regime detection is a decent example, or ranking a basket of symbols. But those are usually a small piece sitting inside a strategy that's still rule-based at its core. Entry logic, exits, sizing — hard to beat simple rules there.
The problem is the failure mode is much worse. A rule-based system that stops working, you can usually look at it and figure out why. A model that stops working, you're basically guessing whether the market changed or you overfit it eight months ago and got lucky since. And with financial data you get one path through history, so cross-validation lies to you in ways it doesn't in other fields.
On the adaptive indicator thing Ryan mentioned — my experience matches. The adaptive versions look better on the chart but by the time the parameter has adjusted, the move it was adjusting to is over. You end up trading the same edge with more moving parts.
My rough take: if you can't state what your system does in a couple of sentences, you've probably fit noise. Doesn't matter whether the fitting was done by you or by gradient descent.
When a ML model goes off track, you can't see the reasoning for those trades. Such a lack of interpretability is a dealbreaker for me.
In financial markets, we're not fully dealing with deterministic outcomes, we're dealing with economics, news, and regime shifts. The more you complicate something sometimes, the weaker the results.
Sure some people have some positive results with a ML model, but a robust strategy for me has transparency, not black boxes, and it simply works.
My initial stage of determining whether or not an indicator is valuable is running it through the Tester to merely visualize it on my desired symbol/chart. As they say, "a picture is worth a thousand words," and a moving chart is more like a moving picture. That works for almost every indicator, e.g., adaptive, repainting, etc. I simply have not seen a readily apparent additional edge in my testing adaptive indicators.
At this point, I'm a bit skeptical of replacing "standard" indicators, that already have an edge over the course a 5 year+ test, with adaptive indicators. I've found that constantly changing the parameters/logic of an indicator in real-time based on very recent live data is somewhat of a crapshoot.
I don't want to confuse adaptive indicators with repainting indicators. Some adaptive indicators can repaint history, but not all adaptive indicators repaint history─the latters simply adapt going forward.For me, I have seen better results with standard Indicator than Adaptive ones. Adaptive Indicators just tend to sound more logical, but when tested on the Tester they seem to fumble after an extended period. I haven't really dedicated much time into testing them but the little one I did doesn't seem to work well with the same filters I used on the standard counterpart.
So maybe they have an edge for a shorter period in higher timeframes.
My initial stage of determining whether or not an indicator is valuable is running it through the Tester to merely visualize it on my desired symbol/chart. As they say, "a picture is worth a thousand words," and a moving chart is more like a moving picture.
Strategy Tester has to be one of the best features in MetaTrader 5. It does almost everything that has to do with data analysis. For mean-reversion strategies, most studies wants you to export symbol data, and analyse if the symbol is simply a mean reverting one (among others), but strategy tester does that, or am I the only one that sees that.
I don't think ML is unavoidable at all. Plenty of people trade rule-based systems that have held up for decades, and most of the ML I see posted here is a linear regression with extra steps.
Where I think it actually helps is when you have a lot of inputs and no clean way to combine them by hand. Regime detection is a decent example, or ranking a basket of symbols. But those are usually a small piece sitting inside a strategy that's still rule-based at its core. Entry logic, exits, sizing — hard to beat simple rules there.
The problem is the failure mode is much worse. A rule-based system that stops working, you can usually look at it and figure out why. A model that stops working, you're basically guessing whether the market changed or you overfit it eight months ago and got lucky since. And with financial data you get one path through history, so cross-validation lies to you in ways it doesn't in other fields.
On the adaptive indicator thing Ryan mentioned — my experience matches. The adaptive versions look better on the chart but by the time the parameter has adjusted, the move it was adjusting to is over. You end up trading the same edge with more moving parts.
My rough take: if you can't state what your system does in a couple of sentences, you've probably fit noise. Doesn't matter whether the fitting was done by you or by gradient descent.
That's basically it. Having a lot of input parameters can be bad for a strategy because each input added is mostly overfitting into that past data, just like the many sentences when explaining what the system does. I also think that generic mode in the strategy tester is something that can also deal with having a lot of inputs to find the "cluster of inputs" that works.
ML, I agree, is good for regime detection. And ranking symbols "faster" than strategy tester would.
- Free trading apps
- Over 8,000 signals for copying
- Economic news for exploring financial markets
You agree to website policy and terms of use
Recently I have seen a lot of traders training more and more ML models everyday, but I am also wondering if this has always been the case for creating trading strategies. I know that machine learning models are good for regime detection, and identifying anomalies (among other things).
So, can these improvements be made using pure rule-based algorithmic approach or is machine learning something that cannot be avoided when creating a truly robust strategy. What are your thoughts based on what you have seen or heard?