I am backtesting a value momentum asset allocation strategy and my in sample period is from 2003 to 2011 and out sample from 2012 to 2019. I am optimising a cutoff for value on in sample to allocate on various asset classes.

We were having a discussion regarding this and someone pointed out that presence of crisis in in sample will skew our optimisation towards it as value was very low(overvalued) during that period and essentially optimisation would be biased towards that(give a higher cutoff). That high cutoff will be less likely to happen again because crisis on that scale hasn't happened after 2008

I have 2 objections to this school of thought:

1) My sample consists of 8 other years than just the crisis year 2008, so would optimisation would that greatly be affected by just one year? 2) Is it right to remove outliers specifically from the in sample period just for sake of consistency? Doesn't that beat the point of incorporating different scenarios in in-sample so our model is ready for it. This way we should remove every outlier then, why just 2008, every month return has fallen by >10% is essentially an outlier.

I would like to know thoughts of this community regarding this

  • 3
    $\begingroup$ I think you should shuffle your in-sample and out-of-sample observations by using a proper technique, like FHS, then don't discard anything but use your many new scenarios to run optimizations. By using a proper ARMA-GARCH model you can keep meaningful properties of your time series (mainly autocorrelation and heteroskedasticity), while random scenarios can make you less dependent on a single past event like 2008 crisis. $\endgroup$ – Lisa Ann Jun 1 '19 at 6:05

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.