def test_alignment(self, env):
items = ('a', 'b')
sids = (1, 2)
dts = env.market_minute_window(
env.open_and_closes.market_open[0],
4,
).values
rp = RollingPanel(2, items, sids, initial_dates=dts[1:-1])
frame = pd.DataFrame(
data=np.arange(4).reshape((2, 2)),
columns=sids,
index=items,
)
nan_arr = np.empty((2, 6))
nan_arr.fill(np.nan)
rp.add_frame(dts[-1], frame)
cur = rp.get_current()
data = np.array(
(((np.nan, np.nan), (0, 1)), ((np.nan, np.nan), (2, 3))), float)
expected = pd.Panel(
data,
major_axis=dts[2:],
minor_axis=sids,
items=items,
)
expected.major_axis = expected.major_axis.tz_localize('utc')
tm.assert_panel_equal(
cur,
expected,
)
rp.extend_back(dts[:-2])
cur = rp.get_current()
data = np.array(
(((np.nan, np.nan), (np.nan, np.nan), (np.nan, np.nan), (0, 1)),
((np.nan, np.nan), (np.nan, np.nan), (np.nan, np.nan), (2, 3))),
float)
expected = pd.Panel(
data,
major_axis=dts,
minor_axis=sids,
items=items,
)
expected.major_axis = expected.major_axis.tz_localize('utc')
tm.assert_panel_equal(
cur,
expected,
)
def test_basics(self, window=10):
items = ['bar', 'baz', 'foo']
minor = ['A', 'B', 'C', 'D']
rp = RollingPanel(window, items, minor, cap_multiple=2)
dates = pd.date_range('2000-01-01', periods=30, tz='utc')
major_deque = deque(maxlen=window)
frames = {}
for i, date in enumerate(dates):
frame = pd.DataFrame(np.random.randn(3, 4), index=items,
columns=minor)
rp.add_frame(date, frame)
frames[date] = frame
major_deque.append(date)
result = rp.get_current()
expected = pd.Panel(frames, items=list(major_deque),
major_axis=items, minor_axis=minor)
tm.assert_panel_equal(result, expected.swapaxes(0, 1))
def test_basics(self):
items = ['foo', 'bar', 'baz']
minor = ['A', 'B', 'C', 'D']
window = 10
rp = RollingPanel(window, items, minor, cap_multiple=2)
dates = pd.date_range('2000-01-01', periods=30, tz='utc')
major_deque = deque()
frames = {}
for i in range(30):
frame = pd.DataFrame(np.random.randn(3, 4),
index=items,
columns=minor)
date = dates[i]
rp.add_frame(date, frame)
frames[date] = frame
major_deque.append(date)
if i >= window:
major_deque.popleft()
result = rp.get_current()
expected = pd.Panel(frames,
items=list(major_deque),
major_axis=items,
minor_axis=minor)
tm.assert_panel_equal(result, expected.swapaxes(0, 1))
def run_history_implementations(option='clever', n=500, change_fields=False,
copy=False, n_items=15, n_minor=20,
change_freq=5, window=100):
items = range(n_items)
minor = range(n_minor)
periods = n
dates = pd.date_range('2000-01-01', periods=periods, tz='utc')
frames = {}
if option == 'clever':
rp = RollingPanel(window, items, minor, cap_multiple=2)
major_deque = deque()
for i in range(periods):
# Add a new and drop an field every change_freq iterations
if change_fields and (i % change_freq) == 0:
minor = minor[1:]
minor.append(minor[-1] + 1)
items = items[1:]
items.append(items[-1] + 1)
dummy = pd.DataFrame(np.random.randn(len(items), len(minor)),
index=items, columns=minor)
frame = dummy * (1 + 0.001 * i)
date = dates[i]
rp.add_frame(date, frame)
frames[date] = frame
major_deque.append(date)
if i >= window:
del frames[major_deque.popleft()]
result = rp.get_current()
if copy:
result = result.copy()
else:
major_deque = deque()
dummy = pd.DataFrame(np.random.randn(len(items), len(minor)),
index=items, columns=minor)
for i in range(periods):
frame = dummy * (1 + 0.001 * i)
date = dates[i]
frames[date] = frame
major_deque.append(date)
if i >= window:
del frames[major_deque.popleft()]
result = pd.Panel(frames, items=list(major_deque),
major_axis=items, minor_axis=minor)
def test_get_current_multiple_call_same_tick(self, env):
"""
In old get_current, each call the get_current would copy the data. Thus
changing that object would have no side effects.
To keep the same api, make sure that the raw option returns a copy too.
"""
def data_id(values):
return values.__array_interface__['data']
items = ('a', 'b')
sids = (1, 2)
dts = env.market_minute_window(
env.open_and_closes.market_open[0],
4,
).values
rp = RollingPanel(2, items, sids, initial_dates=dts[1:-1])
frame = pd.DataFrame(
data=np.arange(4).reshape((2, 2)),
columns=sids,
index=items,
)
nan_arr = np.empty((2, 6))
nan_arr.fill(np.nan)
rp.add_frame(dts[-1], frame)
# each get_current call makea a copy
cur = rp.get_current()
cur2 = rp.get_current()
assert data_id(cur.values) != data_id(cur2.values)
# make sure raw follow same logic
raw = rp.get_current(raw=True)
raw2 = rp.get_current(raw=True)
assert data_id(raw) != data_id(raw2)
def test_get_current_multiple_call_same_tick(self):
"""
In old get_current, each call the get_current would copy the data. Thus
changing that object would have no side effects.
To keep the same api, make sure that the raw option returns a copy too.
"""
def data_id(values):
return values.__array_interface__['data']
items = ('a', 'b')
sids = (1, 2)
dts = self.env.market_minute_window(
self.env.open_and_closes.market_open[0], 4,
).values
rp = RollingPanel(2, items, sids, initial_dates=dts[1:-1])
frame = pd.DataFrame(
data=np.arange(4).reshape((2, 2)),
columns=sids,
index=items,
)
nan_arr = np.empty((2, 6))
nan_arr.fill(np.nan)
rp.add_frame(dts[-1], frame)
# each get_current call makea a copy
cur = rp.get_current()
cur2 = rp.get_current()
assert data_id(cur.values) != data_id(cur2.values)
# make sure raw follow same logic
raw = rp.get_current(raw=True)
raw2 = rp.get_current(raw=True)
assert data_id(raw) != data_id(raw2)
def f(option='clever', n=500, copy=False):
items = range(5)
minor = range(20)
window = 100
periods = n
dates = pd.date_range('2000-01-01', periods=periods, tz='utc')
frames = {}
if option == 'clever':
rp = RollingPanel(window, items, minor, cap_multiple=2)
major_deque = deque()
dummy = pd.DataFrame(np.random.randn(len(items), len(minor)),
index=items,
columns=minor)
for i in range(periods):
frame = dummy * (1 + 0.001 * i)
date = dates[i]
rp.add_frame(date, frame)
frames[date] = frame
major_deque.append(date)
if i >= window:
del frames[major_deque.popleft()]
result = rp.get_current()
if copy:
result = result.copy()
else:
major_deque = deque()
dummy = pd.DataFrame(np.random.randn(len(items), len(minor)),
index=items,
columns=minor)
for i in range(periods):
frame = dummy * (1 + 0.001 * i)
date = dates[i]
frames[date] = frame
major_deque.append(date)
if i >= window:
del frames[major_deque.popleft()]
result = pd.Panel(frames,
items=list(major_deque),
major_axis=items,
minor_axis=minor)
def f(option='clever', n=500, copy=False):
items = range(5)
minor = range(20)
window = 100
periods = n
dates = pd.date_range('2000-01-01', periods=periods, tz='utc')
frames = {}
if option == 'clever':
rp = RollingPanel(window, items, minor, cap_multiple=2)
major_deque = deque()
dummy = pd.DataFrame(np.random.randn(len(items), len(minor)),
index=items, columns=minor)
for i in range(periods):
frame = dummy * (1 + 0.001 * i)
date = dates[i]
rp.add_frame(date, frame)
frames[date] = frame
major_deque.append(date)
if i >= window:
del frames[major_deque.popleft()]
result = rp.get_current()
if copy:
result = result.copy()
else:
major_deque = deque()
dummy = pd.DataFrame(np.random.randn(len(items), len(minor)),
index=items, columns=minor)
for i in range(periods):
frame = dummy * (1 + 0.001 * i)
date = dates[i]
frames[date] = frame
major_deque.append(date)
if i >= window:
del frames[major_deque.popleft()]
result = pd.Panel(frames, items=list(major_deque),
major_axis=items, minor_axis=minor)
def test_adding_and_dropping_items(self, n_items=5, n_minor=10, window=10,
periods=30):
np.random.seed(123)
items = deque(range(n_items))
minor = deque(range(n_minor))
expected_items = deque(range(n_items))
expected_minor = deque(range(n_minor))
first_non_existant = max(n_items, n_minor) + 1
# We want to add new columns with random order
add_items = np.arange(first_non_existant, first_non_existant + periods)
np.random.shuffle(add_items)
rp = RollingPanel(window, items, minor, cap_multiple=2)
dates = pd.date_range('2000-01-01', periods=periods, tz='utc')
frames = {}
expected_frames = deque(maxlen=window)
expected_dates = deque()
for i, (date, add_item) in enumerate(zip(dates, add_items)):
frame = pd.DataFrame(np.random.randn(n_items, n_minor),
index=items, columns=minor)
if i >= window:
# Old labels and dates should start to get dropped at every
# call
del frames[expected_dates.popleft()]
expected_minor.popleft()
expected_items.popleft()
expected_frames.append(frame)
expected_dates.append(date)
rp.add_frame(date, frame)
frames[date] = frame
result = rp.get_current()
np.testing.assert_array_equal(sorted(result.minor_axis.values),
sorted(expected_minor))
np.testing.assert_array_equal(sorted(result.items.values),
sorted(expected_items))
tm.assert_frame_equal(frame.T,
result.ix[frame.index, -1, frame.columns])
expected_result = pd.Panel(frames).swapaxes(0, 1)
tm.assert_panel_equal(expected_result,
result)
# Insert new items
minor.popleft()
minor.append(add_item)
items.popleft()
items.append(add_item)
expected_minor.append(add_item)
expected_items.append(add_item)
class BatchTransform(object):
"""Base class for batch transforms with a trailing window of
variable length. As opposed to pure EventWindows that get a stream
of events and are bound to a single SID, this class creates stream
of pandas DataFrames with each colum representing a sid.
There are two ways to create a new batch window:
(i) Inherit from BatchTransform and overload get_value(data).
E.g.:
```
class MyBatchTransform(BatchTransform):
def get_value(self, data):
# compute difference between the means of sid 0 and sid 1
return data[0].mean() - data[1].mean()
```
(ii) Use the batch_transform decorator.
E.g.:
```
@batch_transform
def my_batch_transform(data):
return data[0].mean() - data[1].mean()
```
In your algorithm you would then have to instantiate
this in the initialize() method:
```
self.my_batch_transform = MyBatchTransform()
```
To then use it, inside of the algorithm handle_data(), call the
handle_data() of the BatchTransform and pass it the current event:
```
result = self.my_batch_transform(data)
```
"""
def __init__(self,
func=None,
refresh_period=0,
window_length=None,
clean_nans=True,
sids=None,
fields=None,
compute_only_full=True,
bars='daily',
downsample=False):
"""Instantiate new batch_transform object.
:Arguments:
func : python function <optional>
If supplied will be called after each refresh_period
with the data panel and all args and kwargs supplied
to the handle_data() call.
refresh_period : int
Interval to wait between advances in the window.
window_length : int
How many days the trailing window should have.
clean_nans : bool <default=True>
Whether to (forward) fill in nans.
sids : list <optional>
Which sids to include in the moving window. If not
supplied sids will be extracted from incoming
events.
fields : list <optional>
Which fields to include in the moving window
(e.g. 'price'). If not supplied, fields will be
extracted from incoming events.
compute_only_full : bool <default=True>
Only call the user-defined function once the window is
full. Returns None if window is not full yet.
downsample : bool <default=False>
If true, downsample bars to daily bars. Otherwise, do nothing.
"""
if func is not None:
self.compute_transform_value = func
else:
self.compute_transform_value = self.get_value
self.clean_nans = clean_nans
self.compute_only_full = compute_only_full
# no need to down sample if the bars are already daily
self.downsample = downsample and (bars == 'minute')
# How many bars are in a day
self.bars = bars
if self.bars == 'daily':
self.bars_in_day = 1
elif self.bars == 'minute':
self.bars_in_day = int(6.5 * 60)
else:
raise ValueError('%s bars not understood.' % self.bars)
# The following logic is to allow pre-specified sid filters
# to operate on the data, but to also allow new symbols to
# enter the batch transform's window IFF a sid filter is not
# specified.
if sids is not None:
if isinstance(sids, (string_types, Integral)):
self.static_sids = set([sids])
else:
self.static_sids = set(sids)
else:
self.static_sids = None
self.initial_field_names = fields
if isinstance(self.initial_field_names, string_types):
self.initial_field_names = [self.initial_field_names]
self.field_names = set()
self.refresh_period = refresh_period
check_window_length(window_length)
self.window_length = window_length
self.trading_days_total = 0
self.window = None
self.full = False
# Set to -inf essentially to cause update on first attempt.
self.last_dt = pd.Timestamp('1900-1-1', tz='UTC')
self.updated = False
self.cached = None
self.last_args = None
self.last_kwargs = None
# Data panel that provides bar information to fill in the window,
# when no bar ticks are available from the data source generator
# Used in universes that 'rollover', e.g. one that has a different
# set of stocks per quarter
self.supplemental_data = None
self.rolling_panel = None
self.daily_rolling_panel = None
def handle_data(self, data, *args, **kwargs):
"""
Point of entry. Process an event frame.
"""
# extract dates
dts = [event.dt for event in itervalues(data._data)]
# we have to provide the event with a dt. This is only for
# checking if the event is outside the window or not so a
# couple of seconds shouldn't matter. We don't add it to
# the data parameter, because it would mix dt with the
# sid keys.
event = Event()
event.dt = max(dts)
event.data = {k: v.__dict__ for k, v in iteritems(data._data)
# Need to check if data has a 'length' to filter
# out sids without trade data available.
# TODO: expose more of 'no trade available'
# functionality to zipline
if len(v)}
# only modify the trailing window if this is
# a new event. This is intended to make handle_data
# idempotent.
if self.last_dt < event.dt:
self.updated = True
self._append_to_window(event)
else:
self.updated = False
# return newly computed or cached value
return self.get_transform_value(*args, **kwargs)
def _init_panels(self, sids):
if self.downsample:
self.rolling_panel = RollingPanel(self.bars_in_day,
self.field_names, sids)
self.daily_rolling_panel = RollingPanel(self.window_length,
self.field_names, sids)
else:
self.rolling_panel = RollingPanel(self.window_length *
self.bars_in_day,
self.field_names, sids)
def _append_to_window(self, event):
self.field_names = self._get_field_names(event)
if self.static_sids is None:
sids = set(event.data.keys())
else:
sids = self.static_sids
# the panel sent to the transform code will have
# columns masked with this set of sids. This is how
# we guarantee that all (and only) the sids sent to the
# algorithm's handle_data and passed to the batch
# transform. See the get_data method to see it applied.
# N.B. that the underlying panel grows monotonically
# if the set of sids changes over time.
self.latest_sids = sids
# Create rolling panel if not existant
if self.rolling_panel is None:
self._init_panels(sids)
# Store event in rolling frame
self.rolling_panel.add_frame(event.dt,
pd.DataFrame(event.data,
index=self.field_names,
columns=sids))
# update trading day counters
# we may get events from non-trading sources which occurr on
# non-trading days. The book-keeping for market close and
# trading day counting should only consider trading days.
if trading.environment.is_trading_day(event.dt):
_, mkt_close = trading.environment.get_open_and_close(event.dt)
if self.bars == 'daily':
# Daily bars have their dt set to midnight.
mkt_close = trading.environment.normalize_date(mkt_close)
if event.dt == mkt_close:
if self.downsample:
downsample_panel(self.rolling_panel,
self.daily_rolling_panel,
mkt_close
)
self.trading_days_total += 1
self.mkt_close = mkt_close
self.last_dt = event.dt
if self.trading_days_total >= self.window_length:
self.full = True
def get_transform_value(self, *args, **kwargs):
"""Call user-defined batch-transform function passing all
arguments.
Note that this will only call the transform if the datapanel
has actually been updated. Otherwise, the previously, cached
value will be returned.
"""
if self.compute_only_full and not self.full:
return None
#################################################
# Determine whether we should call the transform
# 0. Support historical/legacy usage of '0' signaling,
# 'update on every bar'
if self.refresh_period == 0:
period_signals_update = True
else:
# 1. Is the refresh period over?
period_signals_update = (
self.trading_days_total % self.refresh_period == 0)
# 2. Have the args or kwargs been changed since last time?
args_updated = args != self.last_args or kwargs != self.last_kwargs
# 3. Is this a downsampled batch, and is the last event mkt close?
downsample_ready = not self.downsample or \
self.last_dt == self.mkt_close
recalculate_needed = downsample_ready and \
(args_updated or (period_signals_update and self.updated))
###################################################
if recalculate_needed:
self.cached = self.compute_transform_value(
self.get_data(),
*args,
**kwargs
)
self.last_args = args
self.last_kwargs = kwargs
return self.cached
def get_data(self):
"""Create a pandas.Panel (i.e. 3d DataFrame) from the
events in the current window.
Returns:
The resulting panel looks like this:
index : field_name (e.g. price)
major axis/rows : dt
minor axis/colums : sid
"""
if self.downsample:
data = self.daily_rolling_panel.get_current()
else:
data = self.rolling_panel.get_current()
if self.supplemental_data is not None:
for item in data.items:
if item not in self.supplemental_data.items:
continue
for dt in data.major_axis:
try:
supplemental_for_dt = self.supplemental_data.ix[
item, dt, :]
except KeyError:
# Only filling in data available in supplemental data.
supplemental_for_dt = None
if supplemental_for_dt is not None:
data[item].ix[dt] = \
supplemental_for_dt.combine_first(
data[item].ix[dt])
# screen out sids no longer in the multiverse
data = data.ix[:, :, self.latest_sids]
if self.clean_nans:
# Fills in gaps of missing data during transform
# of multiple stocks. E.g. we may be missing
# minute data because of illiquidity of one stock
data = data.fillna(method='ffill')
# Hold on to a reference to the data,
# so that it's easier to find the current data when stepping
# through with a debugger
self._curr_data = data
return data
def get_value(self, *args, **kwargs):
raise NotImplementedError(
"Either overwrite get_value or provide a func argument.")
def __call__(self, f):
self.compute_transform_value = f
return self.handle_data
def _extract_field_names(self, event):
# extract field names from sids (price, volume etc), make sure
# every sid has the same fields.
sid_keys = []
for sid in itervalues(event.data):
keys = set([name for name, value in sid.items()
if isinstance(value,
(int,
float,
numpy.integer,
numpy.float,
numpy.long))
])
sid_keys.append(keys)
# with CUSTOM data events, there may be different fields
# per sid. So the allowable keys are the union of all events.
union = set.union(*sid_keys)
unwanted_fields = set(['portfolio', 'sid', 'dt', 'type', 'source_id'])
return union - unwanted_fields
def _get_field_names(self, event):
if self.initial_field_names is not None:
return self.initial_field_names
else:
self.latest_names = self._extract_field_names(event)
return set.union(self.field_names, self.latest_names)
def test_alignment(self, env):
items = ('a', 'b')
sids = (1, 2)
dts = env.market_minute_window(
env.open_and_closes.market_open[0], 4,
).values
rp = RollingPanel(2, items, sids, initial_dates=dts[1:-1])
frame = pd.DataFrame(
data=np.arange(4).reshape((2, 2)),
columns=sids,
index=items,
)
nan_arr = np.empty((2, 6))
nan_arr.fill(np.nan)
rp.add_frame(dts[-1], frame)
cur = rp.get_current()
data = np.array((((np.nan, np.nan),
(0, 1)),
((np.nan, np.nan),
(2, 3))),
float)
expected = pd.Panel(
data,
major_axis=dts[2:],
minor_axis=sids,
items=items,
)
expected.major_axis = expected.major_axis.tz_localize('utc')
tm.assert_panel_equal(
cur,
expected,
)
rp.extend_back(dts[:-2])
cur = rp.get_current()
data = np.array((((np.nan, np.nan),
(np.nan, np.nan),
(np.nan, np.nan),
(0, 1)),
((np.nan, np.nan),
(np.nan, np.nan),
(np.nan, np.nan),
(2, 3))),
float)
expected = pd.Panel(
data,
major_axis=dts,
minor_axis=sids,
items=items,
)
expected.major_axis = expected.major_axis.tz_localize('utc')
tm.assert_panel_equal(
cur,
expected,
)
class BatchTransform(object):
"""Base class for batch transforms with a trailing window of
variable length. As opposed to pure EventWindows that get a stream
of events and are bound to a single SID, this class creates stream
of pandas DataFrames with each colum representing a sid.
There are two ways to create a new batch window:
(i) Inherit from BatchTransform and overload get_value(data).
E.g.:
```
class MyBatchTransform(BatchTransform):
def get_value(self, data):
# compute difference between the means of sid 0 and sid 1
return data[0].mean() - data[1].mean()
```
(ii) Use the batch_transform decorator.
E.g.:
```
@batch_transform
def my_batch_transform(data):
return data[0].mean() - data[1].mean()
```
In your algorithm you would then have to instantiate
this in the initialize() method:
```
self.my_batch_transform = MyBatchTransform()
```
To then use it, inside of the algorithm handle_data(), call the
handle_data() of the BatchTransform and pass it the current event:
```
result = self.my_batch_transform(data)
```
"""
def __init__(self,
func=None,
refresh_period=0,
window_length=None,
clean_nans=True,
sids=None,
fields=None,
compute_only_full=True,
bars='daily',
downsample=False):
"""Instantiate new batch_transform object.
:Arguments:
func : python function <optional>
If supplied will be called after each refresh_period
with the data panel and all args and kwargs supplied
to the handle_data() call.
refresh_period : int
Interval to wait between advances in the window.
window_length : int
How many days the trailing window should have.
clean_nans : bool <default=True>
Whether to (forward) fill in nans.
sids : list <optional>
Which sids to include in the moving window. If not
supplied sids will be extracted from incoming
events.
fields : list <optional>
Which fields to include in the moving window
(e.g. 'price'). If not supplied, fields will be
extracted from incoming events.
compute_only_full : bool <default=True>
Only call the user-defined function once the window is
full. Returns None if window is not full yet.
downsample : bool <default=False>
If true, downsample bars to daily bars. Otherwise, do nothing.
"""
if func is not None:
self.compute_transform_value = func
else:
self.compute_transform_value = self.get_value
self.clean_nans = clean_nans
self.compute_only_full = compute_only_full
# no need to down sample if the bars are already daily
self.downsample = downsample and (bars == 'minute')
# How many bars are in a day
self.bars = bars
if self.bars == 'daily':
self.bars_in_day = 1
elif self.bars == 'minute':
self.bars_in_day = int(6.5 * 60)
else:
raise ValueError('%s bars not understood.' % self.bars)
# The following logic is to allow pre-specified sid filters
# to operate on the data, but to also allow new symbols to
# enter the batch transform's window IFF a sid filter is not
# specified.
if sids is not None:
if isinstance(sids, (basestring, Integral)):
self.static_sids = set([sids])
else:
self.static_sids = set(sids)
else:
self.static_sids = None
self.initial_field_names = fields
if isinstance(self.initial_field_names, basestring):
self.initial_field_names = [self.initial_field_names]
self.field_names = set()
self.refresh_period = refresh_period
check_window_length(window_length)
self.window_length = window_length
self.trading_days_total = 0
self.window = None
self.full = False
# Set to -inf essentially to cause update on first attempt.
self.last_dt = pd.Timestamp('1900-1-1', tz='UTC')
self.updated = False
self.cached = None
self.last_args = None
self.last_kwargs = None
# Data panel that provides bar information to fill in the window,
# when no bar ticks are available from the data source generator
# Used in universes that 'rollover', e.g. one that has a different
# set of stocks per quarter
self.supplemental_data = None
self.rolling_panel = None
self.daily_rolling_panel = None
def handle_data(self, data, *args, **kwargs):
"""
Point of entry. Process an event frame.
"""
# extract dates
dts = [event.datetime for event in data.itervalues()]
# we have to provide the event with a dt. This is only for
# checking if the event is outside the window or not so a
# couple of seconds shouldn't matter. We don't add it to
# the data parameter, because it would mix dt with the
# sid keys.
event = Event()
event.dt = max(dts)
event.data = {k: v.__dict__ for k, v in data.iteritems()
# Need to check if data has a 'length' to filter
# out sids without trade data available.
# TODO: expose more of 'no trade available'
# functionality to zipline
if len(v)}
# only modify the trailing window if this is
# a new event. This is intended to make handle_data
# idempotent.
if self.last_dt < event.dt:
self.updated = True
self._append_to_window(event)
else:
self.updated = False
# return newly computed or cached value
return self.get_transform_value(*args, **kwargs)
def _init_panels(self, sids):
if self.downsample:
self.rolling_panel = RollingPanel(self.bars_in_day,
self.field_names, sids)
self.daily_rolling_panel = RollingPanel(self.window_length,
self.field_names, sids)
else:
self.rolling_panel = RollingPanel(self.window_length *
self.bars_in_day,
self.field_names, sids)
def _append_to_window(self, event):
self.field_names = self._get_field_names(event)
if self.static_sids is None:
sids = set(event.data.keys())
else:
sids = self.static_sids
# the panel sent to the transform code will have
# columns masked with this set of sids. This is how
# we guarantee that all (and only) the sids sent to the
# algorithm's handle_data and passed to the batch
# transform. See the get_data method to see it applied.
# N.B. that the underlying panel grows monotonically
# if the set of sids changes over time.
self.latest_sids = sids
# Create rolling panel if not existant
if self.rolling_panel is None:
self._init_panels(sids)
# Store event in rolling frame
self.rolling_panel.add_frame(event.dt,
pd.DataFrame(event.data,
index=self.field_names,
columns=sids))
# update trading day counters
# we may get events from non-trading sources which occurr on
# non-trading days. The book-keeping for market close and
# trading day counting should only consider trading days.
if trading.environment.is_trading_day(event.dt):
_, mkt_close = trading.environment.get_open_and_close(event.dt)
if self.bars == 'daily':
# Daily bars have their dt set to midnight.
mkt_close = trading.environment.normalize_date(mkt_close)
if event.dt == mkt_close:
if self.downsample:
downsample_panel(self.rolling_panel,
self.daily_rolling_panel,
mkt_close
)
self.trading_days_total += 1
self.mkt_close = mkt_close
self.last_dt = event.dt
if self.trading_days_total >= self.window_length:
self.full = True
def get_transform_value(self, *args, **kwargs):
"""Call user-defined batch-transform function passing all
arguments.
Note that this will only call the transform if the datapanel
has actually been updated. Otherwise, the previously, cached
value will be returned.
"""
if self.compute_only_full and not self.full:
return None
#################################################
# Determine whether we should call the transform
# 0. Support historical/legacy usage of '0' signaling,
# 'update on every bar'
if self.refresh_period == 0:
period_signals_update = True
else:
# 1. Is the refresh period over?
period_signals_update = (
self.trading_days_total % self.refresh_period == 0)
# 2. Have the args or kwargs been changed since last time?
args_updated = args != self.last_args or kwargs != self.last_kwargs
# 3. Is this a downsampled batch, and is the last event mkt close?
downsample_ready = not self.downsample or \
self.last_dt == self.mkt_close
recalculate_needed = downsample_ready and \
(args_updated or (period_signals_update and self.updated))
###################################################
if recalculate_needed:
self.cached = self.compute_transform_value(
self.get_data(),
*args,
**kwargs
)
self.last_args = args
self.last_kwargs = kwargs
return self.cached
def get_data(self):
"""Create a pandas.Panel (i.e. 3d DataFrame) from the
events in the current window.
Returns:
The resulting panel looks like this:
index : field_name (e.g. price)
major axis/rows : dt
minor axis/colums : sid
"""
if self.downsample:
data = self.daily_rolling_panel.get_current()
else:
data = self.rolling_panel.get_current()
if self.supplemental_data:
for item in data.items:
if item not in self.supplemental_data.items:
continue
for dt in data.major_axis:
try:
supplemental_for_dt = self.supplemental_data.ix[
item, dt, :]
except KeyError:
# Only filling in data available in supplemental data.
supplemental_for_dt = None
if supplemental_for_dt is not None:
data[item].ix[dt] = \
supplemental_for_dt.combine_first(
data[item].ix[dt])
# screen out sids no longer in the multiverse
data = data.ix[:, :, self.latest_sids]
if self.clean_nans:
# Fills in gaps of missing data during transform
# of multiple stocks. E.g. we may be missing
# minute data because of illiquidity of one stock
data = data.fillna(method='ffill')
# Hold on to a reference to the data,
# so that it's easier to find the current data when stepping
# through with a debugger
self._curr_data = data
return data
def get_value(self, *args, **kwargs):
raise NotImplementedError(
"Either overwrite get_value or provide a func argument.")
def __call__(self, f):
self.compute_transform_value = f
return self.handle_data
def _extract_field_names(self, event):
# extract field names from sids (price, volume etc), make sure
# every sid has the same fields.
sid_keys = []
for sid in event.data.itervalues():
keys = set([name for name, value in sid.items()
if isinstance(value,
(int,
float,
numpy.integer,
numpy.float,
numpy.long))
])
sid_keys.append(keys)
# with CUSTOM data events, there may be different fields
# per sid. So the allowable keys are the union of all events.
union = set.union(*sid_keys)
unwanted_fields = set(['portfolio', 'sid', 'dt', 'type',
'datetime', 'source_id'])
return union - unwanted_fields
def _get_field_names(self, event):
if self.initial_field_names is not None:
return self.initial_field_names
else:
self.latest_names = self._extract_field_names(event)
return set.union(self.field_names, self.latest_names)
def test_adding_and_dropping_items(self,
n_items=5,
n_minor=10,
window=10,
periods=30):
np.random.seed(123)
items = deque(range(n_items))
minor = deque(range(n_minor))
expected_items = deque(range(n_items))
expected_minor = deque(range(n_minor))
first_non_existant = max(n_items, n_minor) + 1
# We want to add new columns with random order
add_items = np.arange(first_non_existant, first_non_existant + periods)
np.random.shuffle(add_items)
rp = RollingPanel(window, items, minor, cap_multiple=2)
dates = pd.date_range('2000-01-01', periods=periods, tz='utc')
frames = {}
expected_frames = deque(maxlen=window)
expected_dates = deque()
for i, (date, add_item) in enumerate(zip(dates, add_items)):
frame = pd.DataFrame(np.random.randn(n_items, n_minor),
index=items,
columns=minor)
if i >= window:
# Old labels and dates should start to get dropped at every
# call
del frames[expected_dates.popleft()]
expected_minor.popleft()
expected_items.popleft()
expected_frames.append(frame)
expected_dates.append(date)
rp.add_frame(date, frame)
frames[date] = frame
result = rp.get_current()
np.testing.assert_array_equal(sorted(result.minor_axis.values),
sorted(expected_minor))
np.testing.assert_array_equal(sorted(result.items.values),
sorted(expected_items))
tm.assert_frame_equal(frame.T, result.ix[frame.index, -1,
frame.columns])
expected_result = pd.Panel(frames).swapaxes(0, 1)
tm.assert_panel_equal(expected_result, result)
# Insert new items
minor.popleft()
minor.append(add_item)
items.popleft()
items.append(add_item)
expected_minor.append(add_item)
expected_items.append(add_item)
def run_history_implementations(option='clever',
n=500,
change_fields=False,
copy=False,
n_items=15,
n_minor=20,
change_freq=5,
window=100):
items = range(n_items)
minor = range(n_minor)
periods = n
dates = pd.date_range('2000-01-01', periods=periods, tz='utc')
frames = {}
if option == 'clever':
rp = RollingPanel(window, items, minor, cap_multiple=2)
major_deque = deque()
for i in range(periods):
# Add a new and drop an field every change_freq iterations
if change_fields and (i % change_freq) == 0:
minor = minor[1:]
minor.append(minor[-1] + 1)
items = items[1:]
items.append(items[-1] + 1)
dummy = pd.DataFrame(np.random.randn(len(items), len(minor)),
index=items,
columns=minor)
frame = dummy * (1 + 0.001 * i)
date = dates[i]
rp.add_frame(date, frame)
frames[date] = frame
major_deque.append(date)
if i >= window:
del frames[major_deque.popleft()]
result = rp.get_current()
if copy:
result = result.copy()
else:
major_deque = deque()
dummy = pd.DataFrame(np.random.randn(len(items), len(minor)),
index=items,
columns=minor)
for i in range(periods):
frame = dummy * (1 + 0.001 * i)
date = dates[i]
frames[date] = frame
major_deque.append(date)
if i >= window:
del frames[major_deque.popleft()]
result = pd.Panel(frames,
items=list(major_deque),
major_axis=items,
minor_axis=minor)