Hello to everyone and thanks again for your help, i have find this forum really helpful while working on my final dissertation.

However I'm here again because I have loads of doubts regarding the in-sample and out of sample test of my model. I will explain from my point of view and I would like to know from you if I am completely wrong or if there is something true in what i believe. So my setup is:

  1. my dataset is made of daily return of the S&P500 from 1st of January 2016 till the 31st December 2018
  2. I have performed my optimization on the first two years 1st January2016 till 31st December 2017 and found the various optimal weights i.e Global mean variance portfolio, tangency portfolio, and few other constrained portfolios.
  3. Now suppose i want to perform an insample test and an out of sample test, both of 1 year period, on this hand i wanted to proceed as follows:
  4. for the in-sample test i use the optimal weights i have found with my sample of 2 years, and I see how these portfolios perform using the return of the year 2017-2018. then I find the expected returns, and the measure of risk from these portfolios observation.
  5. for the out of sample instead, the method is to use 1 year of observation as starting point the year 2017-2018 and moving with a rolling window rebalancing my portfolio every day for 1 year (i assume no transaction cost)

Now I am wondering if this could make any sense, every help and suggestion will be accepted. Ps i know that i don't have enough data to perform the optimization, but for the moment i can't change my dataset.

  • $\begingroup$ yes, thank you i edited the question, but i meant two years $\endgroup$ – renato Jun 7 '19 at 19:24
  • $\begingroup$ at point 3 you mean Dec 2017 - Dec 2018 and at point 4 Dec 2017 - Dec 2018? $\endgroup$ – Vitomir Jun 8 '19 at 9:52
  • $\begingroup$ yes Vitomir, do you think does it make sense any sense? if not, what could be an alternative in order to test my restults? $\endgroup$ – renato Jun 8 '19 at 12:46
  • 1
    $\begingroup$ Seems to make sense, I would add cross-validation for in-sample. As you already pointed out, re-balancing daily with no transaction costs can hide the real value of the strategy, especially with regards to the liquidity of the various components. I also understand that by rolling out-of-sample you mean that each day you calculate new returns and covariance matrix, right? In this case seems alright to me $\endgroup$ – Vitomir Jun 8 '19 at 18:02
  • 1
    $\begingroup$ I think you have the right approach. However, you may get some negative feedback if you rebalance daily with no transaction costs because your results are not achievable in the real world (and it kind of looks lazy). Maybe consider adding in transaction costs or rebalancing on a different timeframe ie weekly, monthly, quarterly, etc. Also, if you are using two years of data for the optimization, i would consider reporting the average annual performance and risk statistics for that entire "in sample" timeframe, instead of just the last year. I think it would be more typical to see it this way. $\endgroup$ – Jared Marks Jun 9 '19 at 0:07

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.