One question on the crisis protection program (but I think it also true for the S&P 500 program too), how do you evaluate that the program is not overfitting the historical data especially when it comes to rare occurrences like a crisis?
1. Mechanics: We always start with clear intuition and mechanical reasons a strategy should work. Crucially— does it load on carry, trend, or reversion? Why should those relationships work in a given market? Is our implantation robust?
2. Continuous vs Binary Signals: All our signals change every single day— and as a result, so do our expected returns. Testing whether forward returns co-move with signal changes increases the sample dramatically.
3. Sampling: Train vs test periods— which can be applied by segmenting your data and/or applying your signals to a totally different asset and seeing if they hold up.
We always apply all three levers while constructing any strategy.
One question on the crisis protection program (but I think it also true for the S&P 500 program too), how do you evaluate that the program is not overfitting the historical data especially when it comes to rare occurrences like a crisis?
There are three things to avoid overfitting:
1. Mechanics: We always start with clear intuition and mechanical reasons a strategy should work. Crucially— does it load on carry, trend, or reversion? Why should those relationships work in a given market? Is our implantation robust?
2. Continuous vs Binary Signals: All our signals change every single day— and as a result, so do our expected returns. Testing whether forward returns co-move with signal changes increases the sample dramatically.
3. Sampling: Train vs test periods— which can be applied by segmenting your data and/or applying your signals to a totally different asset and seeing if they hold up.
We always apply all three levers while constructing any strategy.
Looking good!
Gotta get that overlay for ETF Portoflio done now 🔥
This new format is great. The composite allocation breakdown is helpful.
Glad to hear it !