How Hedge Funds Use AI to Develop Trading Strategies.
Caveat. There are ~ 10k hedge funds of all sizes and styles. What I'm demonstrating is how algorithmic hf's (some) that I know use AI to develop -- and put into production, trading strategies. No names or specific HF strategies will be shown.
The example I've built is on BTC daily. I only used three indicators from the basic technical section of available indicators. The idea is to show that you can use basic indicators to build robust (extremely important) strategies that give an edge to the trader. Most people understand that taken on their own, MA, RSI, MACD etc. are unprofitable. When they are incorporated into an optimized strategy, they will contribute to a strategy that makes excellent returns, Sharpe Ratio etc -- yet they never hold up and immediately fail in OOS (out of sample) or production. Why is this? There are many reasons, but one of the most prominent is that you've modeled noise. This causes the strategy to be overfit and non-robust. How do we get around this? It's an optimization problem. As I'll point out in the three indicators used to make this sample system, they were individually optimized using a different technique and AI.
Instead of optimizing a strategy for a specific profit objective, we optimize the underlying indicators individually to a "perfect signal". This modeling is done by applying statistical correlation to the perfect signal which reveals the true settings of the indicator to make a robust strategy.
What is a "perfect signal" and how is it used? There are several techniques depending on what you're modeling and or trying to achieve. The technique used on the three indicators on this chart is called SSA (singular spectrum analysis). Those interested in the technique should Google it because it's outside the scope of these comments. There are others as well such as Wavelet Scattering for feature extraction etc. Ok, but how do we use it?
1) Isolate the indicator you want to optimize.
2) Then run a statistical correlation whereby the SSA is applied to the underlying indicator. The statistical factors are.
a) average error
b) correlation (r)
c) mean squared error
c) % correct
3) Optimize to either
a) minimize error
or
b) maximize correlation
This correlation exercise uses a neural network, one bar look forward to achieve the optimal result. The indicator being optimized uses close or the output of another indicator.
What we are attempting is to create a zero lag, synthetic time series represented by the indicator chosen. This "corrected signal" is now the input to the trading strategy. Because we've used three indicators in this example, they must be synthesized into one cohesive/tradable signal. I won't demonstrate that completely in this example, however, by simple visual alignment -- you can see the trading opportunities.
The indicators.
MA Ribbon. The optimization shows that the true signals in the ribbon are not 20,50,100,200. They are actually 440,497,685,950. This reveals a completely different/true trend structure.
NOTE. Usually you would never think to optimize anything related to OHLC, moving averages etc. Why? Because if the raw price exceeds any look back in your model, the model will break. Also, modeling/predicting the close is too noisy due to degree of freedom and will result in non-robust strategies. The proper technique is to normalize the data to a 0-100 or 100-0-100 range. This is what I did to the MA Ribbon. I then backed out the optimization calculation to periods on the actual price panel. This is the correct technique to do so.
MACD. The MACD signal now more correctly correlates with the true SSA signal. Often, the period settings are far higher than standard, yet attempting to capture perfect maxima/minima. -- this is the zero lag, noise correction component.
RSI. A much higher period setting was found. When RSI is used this way, trend data is also included in the output. NOTE. If you are unable to analyze the RSI indicator with advanced techniques -- there are only two settings for raw RSI. They are 2 and 50. 2 gives you a sharp decision making signal and 50 detrends the indicator to give you a trend component. Plot RSI twice with 2 and 50 periods and visualize it for yourself.
-----
An example of how misleading indicators can be in standard setting.
RSI 14 periods.
Non optimized correlation assessment to close.
Start 3/2/2018
End 7/27/2023
1384 Bars.
Avg. Error = 13,800
Correlation (r) = 0.137
R-Squared = 0.0187
Mean Squared Error = 2.652
Perfect correlation = 1. Statistically significant for trading is ~> .7
-----
As you can see, RSI at 14 periods is on avg. $13,800 away from close in error! Unfortunately, you will not make money using RSI at 14 periods.
This is a window into hedge fund algorithmic trading model construction. Again, I used simple indicators to illustrate how, if used correctly, you can actually trade profitably with them. As with all models, markets are stochastic so the setting will need to be updated in the future. Further, this technique can be used for intraday as well.
If enough people are interested, I can recreate this optimized strategy with TSLA, CL, EUR/USD.
Hope it helps!
The example I've built is on BTC daily. I only used three indicators from the basic technical section of available indicators. The idea is to show that you can use basic indicators to build robust (extremely important) strategies that give an edge to the trader. Most people understand that taken on their own, MA, RSI, MACD etc. are unprofitable. When they are incorporated into an optimized strategy, they will contribute to a strategy that makes excellent returns, Sharpe Ratio etc -- yet they never hold up and immediately fail in OOS (out of sample) or production. Why is this? There are many reasons, but one of the most prominent is that you've modeled noise. This causes the strategy to be overfit and non-robust. How do we get around this? It's an optimization problem. As I'll point out in the three indicators used to make this sample system, they were individually optimized using a different technique and AI.
Instead of optimizing a strategy for a specific profit objective, we optimize the underlying indicators individually to a "perfect signal". This modeling is done by applying statistical correlation to the perfect signal which reveals the true settings of the indicator to make a robust strategy.
What is a "perfect signal" and how is it used? There are several techniques depending on what you're modeling and or trying to achieve. The technique used on the three indicators on this chart is called SSA (singular spectrum analysis). Those interested in the technique should Google it because it's outside the scope of these comments. There are others as well such as Wavelet Scattering for feature extraction etc. Ok, but how do we use it?
1) Isolate the indicator you want to optimize.
2) Then run a statistical correlation whereby the SSA is applied to the underlying indicator. The statistical factors are.
a) average error
b) correlation (r)
c) mean squared error
c) % correct
3) Optimize to either
a) minimize error
or
b) maximize correlation
This correlation exercise uses a neural network, one bar look forward to achieve the optimal result. The indicator being optimized uses close or the output of another indicator.
What we are attempting is to create a zero lag, synthetic time series represented by the indicator chosen. This "corrected signal" is now the input to the trading strategy. Because we've used three indicators in this example, they must be synthesized into one cohesive/tradable signal. I won't demonstrate that completely in this example, however, by simple visual alignment -- you can see the trading opportunities.
The indicators.
MA Ribbon. The optimization shows that the true signals in the ribbon are not 20,50,100,200. They are actually 440,497,685,950. This reveals a completely different/true trend structure.
NOTE. Usually you would never think to optimize anything related to OHLC, moving averages etc. Why? Because if the raw price exceeds any look back in your model, the model will break. Also, modeling/predicting the close is too noisy due to degree of freedom and will result in non-robust strategies. The proper technique is to normalize the data to a 0-100 or 100-0-100 range. This is what I did to the MA Ribbon. I then backed out the optimization calculation to periods on the actual price panel. This is the correct technique to do so.
MACD. The MACD signal now more correctly correlates with the true SSA signal. Often, the period settings are far higher than standard, yet attempting to capture perfect maxima/minima. -- this is the zero lag, noise correction component.
RSI. A much higher period setting was found. When RSI is used this way, trend data is also included in the output. NOTE. If you are unable to analyze the RSI indicator with advanced techniques -- there are only two settings for raw RSI. They are 2 and 50. 2 gives you a sharp decision making signal and 50 detrends the indicator to give you a trend component. Plot RSI twice with 2 and 50 periods and visualize it for yourself.
-----
An example of how misleading indicators can be in standard setting.
RSI 14 periods.
Non optimized correlation assessment to close.
Start 3/2/2018
End 7/27/2023
1384 Bars.
Avg. Error = 13,800
Correlation (r) = 0.137
R-Squared = 0.0187
Mean Squared Error = 2.652
Perfect correlation = 1. Statistically significant for trading is ~> .7
-----
As you can see, RSI at 14 periods is on avg. $13,800 away from close in error! Unfortunately, you will not make money using RSI at 14 periods.
This is a window into hedge fund algorithmic trading model construction. Again, I used simple indicators to illustrate how, if used correctly, you can actually trade profitably with them. As with all models, markets are stochastic so the setting will need to be updated in the future. Further, this technique can be used for intraday as well.
If enough people are interested, I can recreate this optimized strategy with TSLA, CL, EUR/USD.
Hope it helps!
Feragatname
Bilgiler ve yayınlar, TradingView tarafından sağlanan veya onaylanan finansal, yatırım, işlem veya diğer türden tavsiye veya tavsiyeler anlamına gelmez ve teşkil etmez. Kullanım Şartları'nda daha fazlasını okuyun.