Cache holidays

[2torus]

Addresses #108
Cache holidays computation in order to improve running time of repeated valid_days calls.

[leonarduschen]

I'm not sure if the computational time bottleneck for valid_days at #108 is really CustomBusinessDay() and not pd.date_range() though. Have you checked @2torus ?

In any case, this looks like an improvement (though perhaps only partially addresses # 108) to me.

[2torus]

@leonarduschen Thanks for the review.

This Dockerfile builds either versions of the package and then runs a small timing script.

I got:

> docker build . -t mcal_old >/dev/null && docker run -it mcal_old
18.30
> docker build --build-arg NEW="true" . -t mcal_new >/dev/null && docker run -it mcal_new 
1.95

My experience is that 90% of the time in valid_days() is spent in holidays() in the current version.

[leonarduschen]

Awesome. I tried caching the whole valid_days() method a few days ago by storing a very long pd.date_range() on class initialization and managed to get similar results, but I decided to scrape the idea off because it feels very inelegant!

This seems like a far superior solution, much cleaner and not much extra memory.

[2torus]

@leonarduschen I agree. The user could ask for 40-50 years worth of dates, you don't want to be taking the caching decision for them. Holidays, otoh, is not something you expect to be changing.

[rsheftel]

Thanks for this.

[2torus]

@leonarduschen @rsheftel Thank you for allowing me to contribute. I would love to help out in the future. Especially with the quantopian merger.

[rsheftel]

@2torus thanks for your work. Maybe a good place to start helping on the merger would be looking at the calendar and code differences and seeing if you can help determine what is the correct calendar days, or help with adding the functionality to trading_calendars to close the gap.

quantopian/trading_calendars#152

quantopian/trading_calendars#151