Hello Alphien,
I had a question about how long the testing period is? For example, is it around 1, 5, 10 year? I ask because some parts of the model I am writing will need to be adjusted for shifts in economic regime, so just wanted to get a sense of the horizon of testing period.
Thanks,
Majid
I meant the length of the testing period that’s outside the sample period 2010-2019 for Lyxor asset allocation alphathon, just to be clear.
Hello Majid,
We do not disclose what the testing period is so we can assess strategy performance on truly out of sample data.
However, if the logic of your strategy works on the sample period without over fitting, it will work on the testing period.
To give you a concrete example, if what you consider a shift in economic regime only happens once or twice over the sample period, the strategy is probably too specific to the sample period and is unlikely to generalise well.
Let me know if you need more information and good luck for the competition !
Aurian
Ok that’s fine. No, the regime I mean would be the same for about a 20-year period, so anything post-2000 would be the same regime. But if the period extends before 2000, then I need to write some more logic and handle some things.
Majid,
I would recommend that your strategy’s logic doesn’t rely on a specific regime.
What I mean by this is:
-it is ok to have a way of identifying regimes and adapt allocations to it
-it is not ok to build a strategy that relies on the fact that a certain regime that existed in the previous 1, 5, 10 or 50 years will extend into the future.
To make it clearer, let’s look at the following examples:
-I identify bearish regimes using some indicator, I have a specific allocation for bearish regimes -> ok
-In the last 20 years I see a pattern of bull markets following crashes. I build a strategy to take advantage of this pattern and assume it will continue into the future -> will not generalise well to out of sample testing and real world implementation.
The goal is to create a strategy that will work in a future that might or might not look like recent or distant past.
If your strategy does that, it will perform well in any possible out of sample testing.
I hope this clarifies it for you, let me know if you need more information.
Aurian