This functionality is a relatively late addition to backtrader and had to be fitted to the already existing internals. This makes it to be not as flexible and 100% feature full as wished, but it can still serve the purpose in many cases.

Although the implementation tried to allow plug and play filter chaining, the pre-existing internals made it difficult to ensure that could always be achieved. As such, some filters may be chained and some others may not.


  • Transform the values provided by a data feed to deliver a different data feed

The implementation was started to simplify the implementation of the two obvious filters which can be directly used via the cerebro API. These are:

  • Resampling (cerebro.resampledata)

    Here the filter transforms the timeframe and compression of the incoming data feed. For example:

    (Seconds, 1) -> (Days, 1)

    That means that the original data feed is delivery bars with a resolution of 1 Second. The Resampling filter intercepts the data and buffers it until it can deliver a 1 Day bar. This will happen when a 1 Second bar from the next day is seen.

  • Replaying (cerebro.replaydata)

    For the same timeframes as above, the filter would use the 1 Second resolution bars to rebuild the 1 Day bar.

    That means that the 1 Day bar is delivered as many times as 1 Second bars are seen, updated to contain the latest information.

    This simulates, for example, how an actual trading day has developed.


    the length of the data, len(data) and therefore the length of the strategy remain unchanged as long as the day doesn’t change.

Filters at work

Given an existing data feed/source you use the addfilter method of the data feed:

data = MyDataFeed(dataname=myname)
data.addfilter(filter, *args, **kwargs)

And even if it happens to be compatible to the resample/replay filter the following can also be done:

data = MyDataFeed(dataname=myname)
data.addfilter(filter, *args, **kwargs)

Filter Interface

A filter must conform to a given interface, being this:

  • A callable which accepts this signature:

    callable(data, *args, **kwargs)


  • A class which can be instantiated and called

    • During instantiation the __init__ method must support the signature:

      def __init__(self, data, *args, **kwargs)
    • The __call__ method bears this signature:

      def __call__(self, data, *args, **kwargs)

      The instance will be called for each new incoming values from the data feed. The *args and *kwargs are the same passed to __init__


      • True: the inner data fetching loop of the data feed must retry fetching data from the feed, becaue the length of the stream was manipulated
      • False even if data may have been edited (example: changed close price), the length of the stream has remain untouched

    In the case of a class based filter 2 additional methods can be implemented

    • last with the following signature:

      def last(self, data, *args, **kwargs)

      This will be called when the data feed is over, allowing the filter to deliver data it may have for example buffered. A typical case is resampling, because a bar is buffered until data from the next time period is seen. When the data feed is over, there is no new data to push the buffered data out.

      last offers the chance to push the buffered data out.


It is obvious that if the filter supports no arguments at all and will be added without any, the signatures can be simplified as in:

def __init__(self, data, *args, **kwargs) -> def __init__(self, data)

A Sample Filter

A very quick filter implementation:

class SessionFilter(object):
    def __init__(self, data):

    def __call__(self, data):
        if data.p.sessionstart <= data.datetime.time() <= data.p.sessionend:
            # bar is in the session
            return False  # tell outer data loop the bar can be processed

        # bar outside of the regular session times
        data.backwards()  # remove bar from data stack
        return True  # tell outer data loop to fetch a new bar

This filter:

  • Uses data.p.sessionstart and data.p.sessionend (standard data feed parameters) to decide if a bar is in the session.

  • If in-the-session the return value is False to indicate nothing was done and the processing of the current bar can continue

  • If not-in-the-session, the bar is removed from the stream and True is returned to indicate a new bar must be fetched.


    the data.backwards() makes uses of the LineBuffer interface. This digs deep into the internals of backtrader.

The use of this filter:

  • Some data feeds contain out of regular trading hours data, which may not be of interest to the trader. With this filter only in-session bars will be considered.

Data Pseudo-API for Filters

In the example above it has been shown how the filter invokes data.backwards() to remove the current bar from the stream. Useful calls from the data feed objects which are meant as a pseudo-API for Filters are:

  • data.backwards(size=1, force=False): removes size bars from the data stream (default is 1) by moving the logical pointer backwards. If force=True, then the physical storage is also removed.

    Removing the physical storage is a delicate operation and is only meant as a hack for internal operations.

  • data.forward(value=float('NaN'), size=1): moves size bars the storage forward, increasing the physical storage if needed be and fills with value

  • data._addtostack(bar, stash=False): adds bar to a stack for later processing. bar is an iterable containing as many values as lines has the data feed.

    If stash=False the bar added to the stack will be consumed immediately by the system at the beginning of the next iteration.

    If stash=True the bar will undergo the entire loop processing including potentially being reparsed by filters

  • data._save2stack(erase=False, force=False): saves the current data bar to the stack for later processing. If erase=True then data.backwards will be invoked and will receive the parameter force

  • data._updatebar(bar, forward=False, ago=0): uses the values in the iterable bar to overwrite the values in the data stream ago positions. With the default ago=0 the current bar will updated. With -1, the previous one.

Another example: Pinkfish Filter

This is an example of a filter that can be chained, and is meant so, to another filter, namely the replay filter. The Pinkfish name is from the library which describes the idea in its main page: using daily data to execute operations which would only be possible with intraday data.

To achieve the effect:

  • A daily bar will be broken in 2 componentes: OHL and then C.

  • Those 2 pieces are chained with replay to have the following happening in the stream:

    With Len X     -> OHL
    With Len X     -> OHLC
    With Len X + 1 -> OHL
    With Len X + 1 -> OHLC
    With Len X + 2 -> OHL
    With Len X + 2 -> OHLC


  • When an OHLC bar is received it is copied into an interable and broken down to become:
    • An OHL bar. Because this concept doesn’t actually exist the closing price is replaced with the opening price to really form an OHLO bar.
    • An C bar whic also doesn’t exist. The reality is that it will be delivered like a tick CCCC
    • The volume if distributed between the 2 parts
    • The current bar is removed from the stream
    • The OHLO part is put onto the stack for immediate processing
    • The CCCC part is put into the stash for processing in the next round
    • Because the stack has something for immediate processing the filter can return False to indicate it.

This filter works together with:

  • The replay filter which puts together the OHLO and CCCC parts to finally deliver an OHLC bar.

The use case:

  • Seeing something like if the maximum today is the highest maximum in the last 20 sessions an issuing a Close order which gets executed with the 2nd tick.

The code:

class DaySplitter_Close(bt.with_metaclass(bt.MetaParams, object)):
    Splits a daily bar in two parts simulating 2 ticks which will be used to
    replay the data:

      - First tick: ``OHLX``

        The ``Close`` will be replaced by the *average* of ``Open``, ``High``
        and ``Low``

        The session opening time is used for this tick


      - Second tick: ``CCCC``

        The ``Close`` price will be used for the four components of the price

        The session closing time is used for this tick

    The volume will be split amongst the 2 ticks using the parameters:

      - ``closevol`` (default: ``0.5``) The value indicate which percentage, in
        absolute terms from 0.0 to 1.0, has to be assigned to the *closing*
        tick. The rest will be assigned to the ``OHLX`` tick.

    **This filter is meant to be used together with** ``cerebro.replaydata``

    params = (
        ('closevol', 0.5),  # 0 -> 1 amount of volume to keep for close

    # replaying = True

    def __init__(self, data):
        self.lastdt = None

    def __call__(self, data):
        # Make a copy of the new bar and remove it from stream
        datadt =  # keep the date

        if self.lastdt == datadt:
            return False  # skip bars that come again in the filter

        self.lastdt = datadt  # keep ref to last seen bar

        # Make a copy of current data for ohlbar
        ohlbar = [data.lines[i][0] for i in range(data.size())]
        closebar = ohlbar[:]  # Make a copy for the close

        # replace close price with o-h-l average
        ohlprice = ohlbar[data.Open] + ohlbar[data.High] + ohlbar[data.Low]
        ohlbar[data.Close] = ohlprice / 3.0

        vol = ohlbar[data.Volume]  # adjust volume
        ohlbar[data.Volume] = vohl = int(vol * (1.0 - self.p.closevol))

        oi = ohlbar[data.OpenInterest]  # adjust open interst
        ohlbar[data.OpenInterest] = 0

        # Adjust times
        dt = datetime.datetime.combine(datadt, data.p.sessionstart)
        ohlbar[data.DateTime] = data.date2num(dt)

        # Ajust closebar to generate a single tick -> close price
        closebar[data.Open] = cprice = closebar[data.Close]
        closebar[data.High] = cprice
        closebar[data.Low] = cprice
        closebar[data.Volume] = vol - vohl
        ohlbar[data.OpenInterest] = oi

        # Adjust times
        dt = datetime.datetime.combine(datadt, data.p.sessionend)
        closebar[data.DateTime] = data.date2num(dt)

        # Update stream
        data.backwards(force=True)  # remove the copied bar from stream
        data._add2stack(ohlbar)  # add ohlbar to stack
        # Add 2nd part to stash to delay processing to next round
        data._add2stack(closebar, stash=True)

        return False  # initial tick can be further processed from stack