I currently have a signal which applies to a large basket of stocks, and I have gotten Alphalens to work as intended for a daily time period.
However, my belief is that this particular signal is more predictive over a week-week timeframe. I have tried to test this by resampling my prices,
prices.resample("W").last(), and likewise with my raw factors (stacking to get it in the right format for AlphaLens)
factors.resample("W").last().stack(). I then want to produce a tearsheet that examines performance 1 week, 1 month, and 1 quarter in advance.
get_clean_factor_and_forward_returns( factor_weekly, prices_weekly, quantiles=5, periods=(1, 4, 12), # 1 week, 1 month, 1 quarter filter_zscore=None)
This runs fine, and seems to confirm my hypothesis that the factor is more predictive on a weekly horizon (higher and more robust IC).
I wonder if anyone could confirm whether I am interacting with AlphaLens in the correct way. I read somewhere that AlphaLens is agnostic to the time-period as long as the factor and price series align (which they do), but one of my causes for concern is the following table:
1D 4D 12D Ann. alpha 1.784 0.337 0.167 beta 0.312 0.259 0.076 Mean Period Wise Return Top Quantile (bps) 42.465 12.156 3.484 Mean Period Wise Return Bottom Quantile (bps) -48.463 -16.652 -9.072 Mean Period Wise Spread (bps) 90.928 28.475 12.336
We can see that the headers are 1D, 4D and 12D. Is this a "typo" from AlphaLens and actually the "D" just means whatever period I'm using?