Advances in financial machine learning (Marcos López de Prado): explanation of snippet 3.1

Question

I have been reading AFML ( Marcos López de Prado ) and I am having trouble understanding snippet 3.1 which provides the following code:

def getDailyVol(close,span0=100):
    # daily vol, reindexed to close
    df0=close.index.searchsorted(close.index-pd.Timedelta(days=1))
    df0=df0[df0>0]
    df0=pd.Series(close.index[df0-1], index=close.index[close.shape[0]-df0.shape[0]:])
    df0=close.loc[df0.index]/close.loc[df0.values].values-1 # daily returns
    df0=df0.ewm(span=span0).std()
    return df0

Could anyone explain what is wrong with computing daily volatility as:

span0 = 100
close.pct_change().ewm(span = span0).std()

The results in the two methods of computation differ as the previous date used in getDailyVol is different from the day before.

For E-Mini S&P 500 Jun 23, I obtain (blue curve obtained using snippet 3.1 vs. orange curve obtained using pct_change):

Could anyone please explain?

Thank you!

Answer 1

I can't comment on why MLDP is calculating returns this way (perhaps there is a reason explained in the book, which I don't have). However, from a pure code point-of-view he seems to be calculating returns (incorrectly) over a 3 day period. For example, for end date 2000-01-05 (Wednesday) he takes the start date 2000-01-03 (Monday).

Take some dummy data:

close = pd.DataFrame([100, 101, 104, 103, 101], index=pd.date_range("2000-01-03", "2000-01-07"))

Now if you print out df0 before he calculates the daily returns, you get the following:

2000-01-05   2000-01-03
2000-01-06   2000-01-04
2000-01-07   2000-01-05

This doesn't look right, the return on 2000-01-05 should just be the value from that day vs the previous day. He calculates 104/100-1 which is 4% vs the actual day-on-day return of 104/101-1 of 2.9%.

I'd stick to using the pandas pct_change(), it's not only a lot less convoluted but also allows easy chaining of commands (e.g. if you want to filter out weekends, lag the series by a day or remove zero-values in your time series).

As a final side note, although I appreciate a lot of his work, his Python snippets should perhaps best be rewritten to match current best practices and code standards.