Boosting secondary model train on residual, Metalabelling secondary model train on modified label. Its different. But you need to take care of overfitting if you implement it that way, becareful.
I got two questions, if anyone can help me. 1.) At around 19:00, what does the meta labelling model take in as input features in this example with Bollinger Band indicator? 2.) with metal labeling, we look for high recall in the primary model, regardless the precision. So, essentially the highest recall I can get is 1, that means I can have a primary model that tells me to buy every time. If that's the case, how is this meta labelling model different from simply using ML to predict price directions?
What is your preferred labelling technique and bar type? Eg. Triple barrier vs. Trend scanning labelling. Volume bars vs Imbalance bars vs Imbalance run bars
@@HudsonThamesResearch thanks for replying! I still don't understand what imbalance or imbalance run bars are trying to use as the event to generate a new candle.. seems difficult to implement and unclear on initial parameters for mlfinlab's imbalance bar function. Keep up the awesome videos! 🇿🇦
So, is this approach correct? First, we train the primary model using historical data. Then, we train a secondary model to identify false positives. For instance, if the primary model predicts a positive outcome, we rely on the secondary model to assess whether to trust that prediction or not. We don't apply this process to negatives since it doesn't impact our pnl. However, considering market stationarity and other factors, I'm unsure how to effectively evaluate these models. Traditional train-test methods are flawed due to the non-stationary nature of the market, and it seems we lack reliable tools to prevent overfitting.
Wonderful presentation. How do I get involved in your open-source project. I'm an ML Engineer who has over ten years experience trading derivatives. Would love to contribute if I could.
For meta labeling, I understand that one could use a primary model to classify a sample as long, short, or neither {1,-1,0}. The primary model would be trained on a training set (t_set 1). Is the secondary model trained on a different training set (t_set 2) that the primary model didn't see? My initial interpretation is that a different training set is used, where the labels are if the primary model predictions are valid bets.
given some primary model trained on a primary data set, the secondary model takes the training set along with the primary model output, which considers direction, as well as additional variables that help discriminate if the primary algorithm is correct. e.g. a trend system with high recall and low accuracy can be complemented by having the secondary model consider trend or funding variables that are not considered in the primary model
Amazing Video. Just a question how would you generate the probability from the meta model ? I know that the final label of the meta model would be the LOGICAL AND between the primary and the secondary model, but how do you get the probability of the bet side ? Thanks
Hello and thank you for posting this. On 28:00 can you explain better point 2 ? do you mean that we should try to predict longer term PnL not daily pnl?
@@HudsonThamesResearch OK thanks again, I might be getting something wrong here. If my strategy trades once a day at NYSE close isn't daily pnl means actually each trade?
I think the theory is amazing. Is it possible to run an example how to achieve this, even with your bollinger band simple strat? The problem many of us have is profitable strats but way too many losing trades. A way to fliter out (not take these trades) would be great. Perhaps a classic ema200 to sell only when price below ema200 may help but to me this is missing out on the big rebound. Pretty please to run an example in code if possible. TA and keep up the great work!!
Really Great Presentation. Thanks for this!
Boosting secondary model train on residual, Metalabelling secondary model train on modified label. Its different. But you need to take care of overfitting if you implement it that way, becareful.
I got two questions, if anyone can help me.
1.) At around 19:00, what does the meta labelling model take in as input features in this example with Bollinger Band indicator?
2.) with metal labeling, we look for high recall in the primary model, regardless the precision. So, essentially the highest recall I can get is 1, that means I can have a primary model that tells me to buy every time. If that's the case, how is this meta labelling model different from simply using ML to predict price directions?
What is your preferred labelling technique and bar type? Eg. Triple barrier vs. Trend scanning labelling. Volume bars vs Imbalance bars vs Imbalance run bars
@@HudsonThamesResearch thanks for replying! I still don't understand what imbalance or imbalance run bars are trying to use as the event to generate a new candle.. seems difficult to implement and unclear on initial parameters for mlfinlab's imbalance bar function. Keep up the awesome videos! 🇿🇦
Thank you
So, is this approach correct? First, we train the primary model using historical data. Then, we train a secondary model to identify false positives. For instance, if the primary model predicts a positive outcome, we rely on the secondary model to assess whether to trust that prediction or not. We don't apply this process to negatives since it doesn't impact our pnl. However, considering market stationarity and other factors, I'm unsure how to effectively evaluate these models. Traditional train-test methods are flawed due to the non-stationary nature of the market, and it seems we lack reliable tools to prevent overfitting.
Wonderful presentation. How do I get involved in your open-source project. I'm an ML Engineer who has over ten years experience trading derivatives. Would love to contribute if I could.
For meta labeling, I understand that one could use a primary model to classify a sample as long, short, or neither {1,-1,0}. The primary model would be trained on a training set (t_set 1).
Is the secondary model trained on a different training set (t_set 2) that the primary model didn't see? My initial interpretation is that a different training set is used, where the labels are if the primary model predictions are valid bets.
given some primary model trained on a primary data set, the secondary model takes the training set along with the primary model output, which considers direction, as well as additional variables that help discriminate if the primary algorithm is correct. e.g. a trend system with high recall and low accuracy can be complemented by having the secondary model consider trend or funding variables that are not considered in the primary model
Amazing Video. Just a question how would you generate the probability from the meta model ? I know that the final label of the meta model would be the LOGICAL AND between the primary and the secondary model, but how do you get the probability of the bet side ? Thanks
What ML / statistic technique was used at 5m00s to cluster the regimes? Is this a useful meta-labeling feature?
Which of the two de Prado books do you think is best to read first? Other than content, how do they differ? Different target audiences?
@@HudsonThamesResearch thanks!
awesome!
Hello and thank you for posting this. On 28:00 can you explain better point 2 ? do you mean that we should try to predict longer term PnL not daily pnl?
@@HudsonThamesResearch I see thanks for your message. So I guess in daily frequency strategy daily pnl will be the way to go.
@@HudsonThamesResearch OK thanks again, I might be getting something wrong here. If my strategy trades once a day at NYSE close isn't daily pnl means actually each trade?
I think the theory is amazing. Is it possible to run an example how to achieve this, even with your bollinger band simple strat? The problem many of us have is profitable strats but way too many losing trades. A way to fliter out (not take these trades) would be great. Perhaps a classic ema200 to sell only when price below ema200 may help but to me this is missing out on the big rebound. Pretty please to run an example in code if possible. TA and keep up the great work!!
Is that a Mzansi accent?