Combining Trading Signals Before Constrained Portfolio Optimization
Summary
The document considers how to combine roughly ten signals across roughly one hundred assets before sending expected returns to a mean-variance portfolio optimizer with transaction costs and constraints. It suggests ridge regression or principal component analysis as ways to form robust linear combinations, while retaining the optimizer to manage portfolio risk, costs, and restrictions.
It also raises a calibration issue: signals may pass through portfolio constraints and cost penalties with different effectiveness, potentially favoring signals that imply large spread positions or leverage. The answer recommends risk adjustment and normalization, attention to scalability and transfer coefficients, and validation through cross-validation and out-of-sample tests across market conditions. These are proposed practices rather than a demonstrated comparison: the document supplies no data, fitted weights, or evidence identifying a best combination method. Signal construction and optimizer behavior therefore need empirical assessment in the specific portfolio setting.
Key ideas
- Ridge regression or principal component analysis can produce linear combinations of multiple trading signals.
- A portfolio optimizer can then apply transaction cost penalties and portfolio constraints.
- Risk adjustment and normalization may help make signals more comparable.
- Different signals may pass through portfolio restrictions and costs with different effectiveness.
- Cross-validation and out-of-sample testing are suggested, but no method is supported by empirical results in the document.
Tags
Full text
# Combining alphas for mean variance optimizer # Combining alphas for mean variance optimizer I have about 10 signals that I want to use to trade about 100 assets. I am using a mean variance optimizer that has a transaction cost penalty and enforces various portfolio constraints. I need to combine my signals in some way to get a combined signal, which I can then pass to my portfolio optimization. One way I can think of is to run some kind of rolling ridge regression of signals vs asset return. Or I could run another mean variance style optimization to get signal weights at each time step. Am I on the right track here? Or are there other ways how this is usually done? The problem is that the restrictions and cost penalty might mean that some signals pass through the optimizer easier than others . Like I would imagine that highly leveraged signals that take large spread positions may have a much higher transfer coefficient than others. I am not sure if there is a good way to control for this or if I even should try to control it ## Answer by Mahavir Bhattacharya (score 1) https://quant.stackexchange.com/a/79907 Your approach seems feasible. Here's what I would consider doing on similar hypotheses: - To combine signals for trading 100 assets, consider ridge regression or PCA to derive robust, linear combinations. - Use mean-variance optimization to manage transaction costs and enforce constraints effectively. - Adjust signals for risk factors and normalize for fair treatment. - Address signal scalability and transfer coefficients to mitigate bias. - Validate methods through cross-validation and robust out-of-sample testing to ensure performance across market conditions. - Iterate based on empirical results for strategy refinement and adaptability in real-world scenarios. Hope it helps!
Shown in full with attribution under the source's licence. Licence: CC BY-SA 4.0 (Stack Exchange)
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.