Begin your Algorithmic Trading journey with the most comprehensive quant trading curriculum with industry experts- "Executive Programme in Algorithmic Trading (EPAT)" Register Now - bit.ly/41mrPqu FREE Ebooks: 🎁Machine Learning In Trading | Step By Step Implementation of ML Models - bit.ly/3SXt7Ww 🎁Algorithmic Trading | A Rought & Ready Guide - bit.ly/48n9b3E 🎁Python Basics | With Illustrations From The Financial Markets - bit.ly/48hqCCW
Hi! Thanks for your comment. Dr Starke covers the complete programming part of it in his lecture in EPAT (The Executive Programme in Algorithmic Trading). If you wish to know more about EPAT, kindly visit: www.quantinsti.com/ Or feel free to connect with us here: bit.ly/2WoWILi
Hello, You can check our course 'Deep Reinforcement Learning in Trading' ( quantra.quantinsti.com/course/deep-reinforcement-learning-trading). The course teaches you to apply reinforcement learning to create, backtest, paper trade, and live trade a strategy using two deep learning neural networks and replay memory.
Great to hear you're interested! 🌟 We have more insightful videos on our channel that you might enjoy. Dive in and discover more about Machine Learning and its applications in Trading in these playlists - bit.ly/3RoPp1t Happy Learning😊
Hello Ashish. Yes, the data should be normalised before passing it as input. You can subtract the mean of data from the data points and divide the result with the standard deviation of data to normalise. Hope this answers your query. 😊
Overfitting is the spectrum of algotrading. It is so difficult to avoid it or compensate it and as it is explained in the video NN will learn noise (because it is dominant in market data) and have no idea of real patterns in the data. If there is any pattern, any memory is so weak that I don't think NN is the way to go at all unless some very clever and radical idea (of how to filter out noise) is applied. I'm not aware of anything like that in the literature so far.
Hello Santostasi. 1) About overfitting: Overfitting is indeed a big issue whenever you optimize the parameters of an ML model. However, there is an increasing consensus that Random Forest algorithms are the best to circumvent this problem, compared to other models. You can suggest to the student to use that model. 2) About the NN model learning from noise: Actually, that happens to any model. Models are applied and they try to find a pattern through the noisy data. Instead of the OHLC data, you can use Tick bars (which have better statistical properties), volume bars, etc. The signal-to-noise ratio in financial markets is really small. It's not so simple to get a signal. Every researcher/trader will face this with any model. The video is focused on the explanation of the model and its application to data. The solution to issues like a low signal-to-noise ratio, finding patterns with a different model, etc can be found in other videos o a future video we'd be more than happy to create. 3) Finding a pattern and memory loss: Price memory is indeed lost in returns. You can use an ARFIMA model applied to the price series and get its residuals. The ARFIMA model is applied as (0,d,0) where d is your control variable: You optimize "d" to get the ARFIMA model residuals to result to be stationary (for example, make the unit root statistic less and close to 5%). These residuals can be used as a prediction feature, instead of simple returns. ARFIMA residuals have the property to be stationary and also maintain the price memory. We hope this helps 🤝
Hi Al, thanks for your comment. You can add Stop Loss and Take Profit to the model. You can define your rewards and penalties accordingly. If a stop-loss is hit, penalise the model, while if a take profit is hit, reward the model. We had tried this approach, but the results were sub-optimal. You can still experiment with this and form your own opinion. You can also try using multiple RL learners and assign them different weights based on a factor of your choice. In case you are still worried about huge losses, you can always manually intervene or hard code a stop-loss and take-profit in your strategy. However, this approach is not recommended. We hope this helps.
Hello Shyam. Stock patterns can be classified using neural networks. The neural networks can be combined with a framework of reinforcement learning. Please refer to the following resources. 1. Stock Pattern Classification from Charts using Deep Learning Algorithms (www.researchgate.net/publication/346543292_Stock_Pattern_Classification_from_Charts_using_Deep_Learning_Algorithms) 2. Reinforcement Learning with Neural Network (www.baeldung.com/cs/reinforcement-learning-neural-network) Thank you
This could be a nice course in RL, great but it is USELESS for trading. The example given in the end is a code that can understand sine waves. Ok, nice (I can do that without using AI). By the time he adds noise algo is already bad, but the fundamental point is markets have no SINE WAVES or any other type of reliable periodicities. If it was that easy one could use Fourier Analysis and pick dominant frequencies in a moving window without using AI. So his work may have interesting applications in other fields but not in trading.
Hello Santostasi. Your point of view is precise, but as we said, we wanted to focus on the explanation of the model and its application. In the Quantra course, you will have an answer about using the model with real-world data! I hope this helps!
Interesting but useless for real trading. There are no patterns, no sine waves, in real market data. Can you make an AI system that deals with real markets?
Hello Santostasi. As we said previously, the video's purpose was the explanation the model and its application to a series. You can try to use the model for real-world data! Or, in case you want great help from our end, you can join the following course quantra.quantinsti.com/course/deep-reinforcement-learning-trading. You will have there the model applied to real-world data!
Begin your Algorithmic Trading journey with the most comprehensive quant trading curriculum with industry experts- "Executive Programme in Algorithmic Trading (EPAT)"
Register Now - bit.ly/41mrPqu
FREE Ebooks:
🎁Machine Learning In Trading | Step By Step Implementation of ML Models - bit.ly/3SXt7Ww
🎁Algorithmic Trading | A Rought & Ready Guide - bit.ly/48n9b3E
🎁Python Basics | With Illustrations From The Financial Markets - bit.ly/48hqCCW
very insightful thank you Dr
This guy has a great voice. Makes learning so calm.
Very well explained. Thank you!
Great video. Thanks
Thanks for your comment. We're glad that you liked it. 🙂
Do subscribe to our channel to always get notified when our latest video goes Live!
Where do we find the reinforcement learning coding video as Dr Thomas Starke mentioned it will be there in Quantintsti lecture?
Hi! Thanks for your comment. Dr Starke covers the complete programming part of it in his lecture in EPAT (The Executive Programme in Algorithmic Trading).
If you wish to know more about EPAT, kindly visit: www.quantinsti.com/
Or feel free to connect with us here: bit.ly/2WoWILi
Thanks
Where can we find a code example of what he describes in this video
Hello,
You can check our course 'Deep Reinforcement Learning in Trading' ( quantra.quantinsti.com/course/deep-reinforcement-learning-trading). The course teaches you to apply reinforcement learning to create, backtest, paper trade, and live trade a strategy using two deep learning neural networks and replay memory.
I am interested in this too
Great to hear you're interested! 🌟 We have more insightful videos on our channel that you might enjoy. Dive in and discover more about Machine Learning and its applications in Trading in these playlists - bit.ly/3RoPp1t
Happy Learning😊
do we need to normalize stock price data before feeding it to reinforcement learning model and if we do how to implement it ?
Hello Ashish. Yes, the data should be normalised before passing it as input. You can subtract the mean of data from the data points and divide the result with the standard deviation of data to normalise.
Hope this answers your query. 😊
Overfitting is the spectrum of algotrading. It is so difficult to avoid it or compensate it and as it is explained in the video NN will learn noise (because it is dominant in market data) and have no idea of real patterns in the data.
If there is any pattern, any memory is so weak that I don't think NN is the way to go at all unless some very clever and radical idea (of how to filter out noise) is applied. I'm not aware of anything like that in the literature so far.
Hello Santostasi.
1) About overfitting: Overfitting is indeed a big issue whenever you optimize the parameters of an ML model. However, there is an increasing consensus that Random Forest algorithms are the best to circumvent this problem, compared to other models. You can suggest to the student to use that model.
2) About the NN model learning from noise: Actually, that happens to any model. Models are applied and they try to find a pattern through the noisy data. Instead of the OHLC data, you can use Tick bars (which have better statistical properties), volume bars, etc. The signal-to-noise ratio in financial markets is really small. It's not so simple to get a signal. Every researcher/trader will face this with any model. The video is focused on the explanation of the model and its application to data. The solution to issues like a low signal-to-noise ratio, finding patterns with a different model, etc can be found in other videos o a future video we'd be more than happy to create.
3) Finding a pattern and memory loss: Price memory is indeed lost in returns. You can use an ARFIMA model applied to the price series and get its residuals. The ARFIMA model is applied as (0,d,0) where d is your control variable: You optimize "d" to get the ARFIMA model residuals to result to be stationary (for example, make the unit root statistic less and close to 5%). These residuals can be used as a prediction feature, instead of simple returns. ARFIMA residuals have the property to be stationary and also maintain the price memory.
We hope this helps 🤝
What about using a fixed stop loss and take profit .. or implementing trailing stop loss.
Hi Al, thanks for your comment. You can add Stop Loss and Take Profit to the model. You can define your rewards and penalties accordingly.
If a stop-loss is hit, penalise the model, while if a take profit is hit, reward the model. We had tried this approach, but the results were sub-optimal. You can still experiment with this and form your own opinion.
You can also try using multiple RL learners and assign them different weights based on a factor of your choice.
In case you are still worried about huge losses, you can always manually intervene or hard code a stop-loss and take-profit in your strategy. However, this approach is not recommended.
We hope this helps.
@@quantinsti Thanks :)
@@alrey72 We're happy to help! Thanks, and stay safe! :)
awesome video !
Thank you for acknowledging.
can we recognize chart patterns through RL ?
Hello Shyam.
Stock patterns can be classified using neural networks. The neural networks can be combined with a framework of reinforcement learning. Please refer to the following resources.
1. Stock Pattern Classification from Charts using Deep Learning Algorithms (www.researchgate.net/publication/346543292_Stock_Pattern_Classification_from_Charts_using_Deep_Learning_Algorithms)
2. Reinforcement Learning with Neural Network (www.baeldung.com/cs/reinforcement-learning-neural-network)
Thank you
1:00
This could be a nice course in RL, great but it is USELESS for trading. The example given in the end is a code that can understand sine waves. Ok, nice (I can do that without using AI). By the time he adds noise algo is already bad, but the fundamental point is markets have no SINE WAVES or any other type of reliable periodicities.
If it was that easy one could use Fourier Analysis and pick dominant frequencies in a moving window without using AI. So his work may have interesting applications in other fields but not in trading.
Hello Santostasi.
Your point of view is precise, but as we said, we wanted to focus on the explanation of the model and its application. In the Quantra course, you will have an answer about using the model with real-world data!
I hope this helps!
Interesting but useless for real trading. There are no patterns, no sine waves, in real market data. Can you make an AI system that deals with real markets?
Hello Santostasi.
As we said previously, the video's purpose was the explanation the model and its application to a series. You can try to use the model for real-world data! Or, in case you want great help from our end, you can join the following course quantra.quantinsti.com/course/deep-reinforcement-learning-trading. You will have there the model applied to real-world data!