This is an optimization where we're building an environment but not
using its finder. Ideally, the consumer would use just the calendar,
but it's not fully featured quite yet.
- Adds `zipline.pipeline.Pipeline`, a new user-facing class for managing
pipelines of Modeling API expressions.
- Adds `attach_pipeline` and `drain_pipeline` as API methods
- Removes `add_factor` and `add_filter` as API methods. These have been
replaced two new methods on `Pipeline`: `add`, and `apply_screen`.
- Adding a `Filter` as a column no longer implicitly truncates rows from
the Modelling API output. It simply causes a new column, of dtype
`bool` to show up in the output. Removal of rows is now handled by the
new `apply_screen` method of `Pipeline`.
- Refactors the existing Modeling API tests to reflect the new APIs.
Overhauls `zipline.modelling.visualize` to use in-memory buffers when
shelling out to `dot` and friends.
Also adds `svg`, `png`, and `jpeg` attributes to `TermGraph`, and adds a
`_repr_png_` so that `TermGraph` renders as a PNG by default.
_convert_asset_str_fields was added to address the fact that the
unicode returned by SQLAlchemy was causing test failures, in
particular test_repr and test_root_symbols failed because of the
leading 'u' character in the unicode of python-2.x. Later it was
pointed out that this method would behave differently in python 2
and 3, in particular the str() method would return bytes in python
2 and unicode in python 3. After investigating it was found that
the _convert_asset_str_fields was no longer necessary. It is unclear
why the original tests were failing. It is not due to SQLAlchemy
versions, as tests pass for version >1 and <1.
- Combine the net value and exposure functions into `calc_net` since
they use the same logic.
- Change the logic to handle on empty list to using the a start value of
0.0. More concise, and reduces the number of return points from the
function to one.
Instead of having two leverage functions, whose differences were the
parameter names, add a `calc_leverage` function, with the calling code
determining whether it is gross or net by the type of exposure passed in.
Instead of calculating the position values for each stat result, e.g.
gross_exposure, net_liquidity etc.; get the positions upfront and then
calculate the period and position stats in order, passing each value
explicitly to the ones that follow it in the dependency chain.
e.g. the gross_value depends on the long_value and the short_value,
which called the position_values property for calculating both the
long_value and the short_value.
Removing the repeated calls to position_values (and
position_exposures) removes the need for the caching the last sale
prices and position amounts in separate vectors, since it is inexpensive
enough to read those values off of the positions dictionary held in the
position tracker.
This patch gives a small gain to ~500 sized portfolios, but the main
intent is to clear the path to not storing last_sale_prices on the
position objects at all. Removing all of the caching layer in this class
makes that change easier to apply. Removing the extra calls to
position_values also made this class easier to step through/reason about
when splicing in the new last sale price access, as well.
Scott Sanderson