Machine Learning, retraining data. Layering models vs new model combined data.

In QuantConnect, you can retrain your models periodically as your algorithm executes. This can be done by calling the Train method as a Scheduled Event in the Initialize method. This allows you to train a new model based on the latest data, which can be useful in real-world trading where market conditions can change over time.

Here is an example of how you can schedule your model to be retrained every Sunday at 8:00 AM:

# Train the model every Sunday at 8:00 AM
self.Train(self.DateRules.Every(DayOfWeek.Sunday), self.TimeRules.At(8, 0), self.my_training_method)

This method can be used with various machine learning libraries supported by QuantConnect, such as Stable Baselines, Aesera, GPlearn, Hmmlearn, Scikit-Learn, Tslearn, and XGBoost.

As for the best approach between layering models to confirm each other's signals, training a new model from the combined data, or both, it can depend on various factors such as the specific trading strategy, the type of data, and the market conditions. It might be beneficial to experiment with different approaches and see which one works best for your specific use case.

You can find more information about training models in QuantConnect in the documentation.

Please note that the effectiveness of these methods can vary and should be thoroughly tested before deploying in a live trading environment.

Sorry, my question isn't how to train models, it's how to properly layer models, model(a) is trained by signal(a), model(b) is trained on only the trades taken by model(a) using signal(a)+model(a) confirming a valid signal using new data that neither model has seen before.

I guess my question is, does it work out better to train  new model from the already filtered data or run models in layers and progressively build up layers over time as the algo runs. 

It's probably a stupid question, layering seems obvious and testing will probably give me an answer but testing is only testing, experience is worth a lot more.

The approach of layering models, where one model's output is used as input for another model, is a common technique in machine learning known as ensemble learning. However, the effectiveness of this approach can vary greatly depending on the specific models and data used.

In the context of trading, using one model to filter trades for another model could potentially improve the performance of your algorithm by reducing the number of false positives. However, it could also lead to overfitting if the models become too complex or too closely tailored to the historical data.

As for whether it's better to train a new model from the already filtered data or to layer models, it can depend on various factors such as the specific trading strategy, the type of data, and the market conditions. It might be beneficial to experiment with both approaches and see which one works best for your specific use case.

In QuantConnect, you can implement this layered approach by using multiple Alpha Models in your algorithm. For example, you could use one Alpha Model to generate initial trade signals, and then use another Alpha Model to filter these signals based on additional criteria. You can determine which Alpha Model's Insights to take inside the Portfolio Construction Model by checking the class of where the Insight came from using Insight.SourceModel.

Unfortunately, the provided extracts do not contain specific code examples for layering models in this way. However, you can find more information about using multiple Alpha Models in the QuantConnect documentation.

Please note that the effectiveness of these methods can vary and should be thoroughly tested before deploying in a live trading environment.