standalone full_handcrafting.py

[littlebigfis]

In Rob's standalone handcrafting script, I see correlation is taken from the last ~100 rows (calc_correlation) but the vol_vector uses the entire dataset (.std() of all weekly returns). Why take full-history volatility instead of a matching short window for everything?

When the script calculates vol from the entire dataset while refreshing correlation on the last ~100 rows, the volatility estimate stays nearly constant for long history instruments (since it’s averaged over the entire history). That means any recent changes in price variation won’t move the vol measure much. As a result, correlation shifts become the main driver for adjusting handcrafting weights. Essentially, the model’s sensitivity to new volatility signals is lost, since the long‐history std is locked in and can overshadow near‐term volatility trends.

[robcarver17]

Because I am optimising things that have the same expected volatility.