The write_data methods invokes the relevant AssetDBWriter subclass
to write data to the database. update_asset_finder is no longer
a relevant method since the AssetFinder is strictly a reader class.
This patch lays the groundwork for a compute engine designed to
facilitate construction of factor-based universe screening and portfolio
allocation. It contains:
A new module, `zipline.modelling`, containing entities that can be used
to express computations as dependency graphs. Each node in such a graph
is an instance of the base `Term` class, defined in
`zipline.modelling.term`. Dependency graphs are executed by instances
of `FFCEngine`, defined in `zipline.modelling.engine`.
A new module, `zipline.data.ffc`, containing loaders and dataset
definitions for inputs to the modelling API.
New `TradingAlgorithm` api methods: `add_factor`, and `add_filter`.
These methods can only be called from `initialize`, and are used to
inform the algorithm that each day it should compute the given terms.
Computed factor results are made available through a new attribute of
the `data` object in `before_trading_start` and `handle_data`. Computed
filter results control which assets are available in the factor matrix
on each day.
The minutely calculation of risk metrics had been removed with a
previous patch, remove vestigial references.
Remove a test which tested the behavior of updating the second minute of
a day.
Remove the logic that changed the datetime index of the risk metrics
depending on emission rate, now only trading_days are needed.
Remove `returns_frequency` parameter since both minute and daily
data frequency always use daily returns.
Attack the startup bottleneck of creating the asset finders caches for a
large universe, which was between 1-2 seconds on development and
production machines.
Instead, allow the AssetFinder to be passed a sqlite3 file that has
already been populated and then hydrate asset objects only when an
equity is referenced for the first time.
To create aforementioned sqlite3, create an AssetFinder with an db_path
and `create_table` set to True. If `create_table` is set to False, the
prepopulated data in the sqlite file found at db_path will be used.
Default behavior is to use an in memory database.
Behavior that changes:
- Fuzzy lookup now only works on one character, that character needs to be
specified at write/metadata consumption time, since the fuzzy lookup key
is created by dropping the character from each symbol.
- Overwriting partially written metadata is no longer
supported. i.e. some unit tests allowed for inserting just the identifier,
and then later updating the symbol, end_date, etc.
Instead of building an upsert behavior at this time, this patch
changes the unit tests so that the data for each asset is only
inserted once.
Other notes:
- populate_cache is now removed, since there is no longer a two step
process of inserting metadata and then realizing that metadata into
assets. _spawn_asset is rolled into insert_metadata, so that a call to
insert_metadata both converts the metadata and makes it available in
the data store.
Instead of using the pandas.Series datetime index for every single
vector, get the index at the beginning of the update loop based on the
dt and then use that index to set the values.
Also, since the dt lookup is no longer needed, store the values as numpy
arrays, which are more lightweight.
Locally, this patch cuts out about 60% of the time spent in the update
method.
- AssetFinder no longer accepts an unused trading_calendar.
- AssetFinder correctly accepts a DataFrame as input.
- Tests for AssetFinder no longer rely on a global trading environment.
Use explicit references to the performance tracker instead of the proxy
lookup.
Mainly a putative change, which surfaced when reasoning about places
where the position tracker and period need access to the last sale
price.
Referencing context.account in handle_data() can block updates to account,
portfolio and performance metrics, which can cause unintended changes in
financial performance. Additionally, since context.account metrics are not
updating frequenctly enough (daily, rather than minutely) algos which base
decisions on these metrics can have undesired behavior.
To address this we do not base our decision to update performance on the
state of account_needs_update.
When we call get_account() from the market and minute close methods, we
set the performance_needs_update parameter to False since we just updated
performance.
_get_account() is renamed to _update_account() since the method does not
return account information.
Also remove redundant asset_needs_update = True lines.
By having both the trade simulation main loop route events to "process"
methods based on event type and the process methods also checking event
type, there was some duplicated effort in doing that comparison many
times.
A particular case where this was noted in profiling was for the
`process_event` function which was checking if the type was not a trade
and returning early, when in a larger universe of stocks the value
returned False 99% of the time.
Instead provide separate process functions specific to each type,
e.g. e.g. `process_trade` and `process_transaction` and route traffic to
those functions in tradesimulation.
For a universe of 160 stocks on both no-op algo and an algo that rebuys
its universe every day, saw about a 10% increase locally.
Also:
- Add process_benchmark to blotter since internal subclass relies on
logic on benchmark, this allows the internal process_trade to be a
`pass`.
- Add warning on unrecoginzed event types.
The risk containers that are actually used for reports use the
'cumulative' style container which has an index of days, not minutes.
The minute containers and copying of data etc. were causing an expanding
memory footprint.
The intraday_risk_metrics is being removed since the values are not
used; cumulative risk metrics with the last value updated to the latest
close has been used for some time.
Before the removal of intraday_risk_metrics, the position trackers
passing of benchmark returns to the cumulative risk metrics needs to no
longer depend on the calculations done by the intraday stats. So instead
use the all_benchmark_returns stored in the tracker directly.
For beta calculation:
Remove `.dropna` , since it was creating a new
Series and Index which inflated memory usage as algorithm run time
progressed.
For downside risk calculations:
Instead of using pd.Series calculations, pass the underlying
numpy array which have already been sliced to the exact dt, so that the
call to `round` does not create a new Series.
Remove use of defaultdict for orders_by_modified, which was causing an
empty list to be added every time to_dict was called with a specified
dt.
Nnoticed in the minute emission case when hunting another memory leak,
every simulation minute a new Timestamp and list was created and never
let go.
Only fill limit order if impacted fill price is better than the limit price.
If a limit order is partially filled, only fill the remaining shares if the
impacted fill price is better than the limit price.
