I usually test new nodes:
- On a portfolio greater than 25 to avoid noise from individual stocks
- Test in 10-20 years (especially to test it in a period like 2008)
- Tests on the universe I’ll be using and with the correct rebalancing period, but also breaks it down into 2-3 subuniverses, and with different starting points.
- Tests against different universes
The original system is a multifactor system with around 80 nodes grouped in all 5 known factors.
I decide if the new node should join if it offers better returns for the system than the original system.
But would it be better to consider whether the node should join by looking at whether it gets better returns in the rolling backtest rather than the simulator? Why, why not?