Hidden Powers of Python (3)

Last, but not least, in this series about how the hidden powers of Python are used in backtrader is how some of the magic variables show up.

Where do self.datas and others come from?

The usual suspect classes (or subclasses thereof) Strategy, Indicator, Analyzer, Observer have auto-magically defined attributes, like for example the array which contains the data feeds.

Data Feeds are added to a cerebro instance like this:

from datetime import datetime
import backtrader as bt

cerebro = bt.Cerebro()
data = bt.YahooFinanceData(dataname=my_ticker, fromdate=datetime(2016, 1, 1))


Our winning strategy for the example will go long when the close goes above a Simple Moving Average. We’ll use Signals to make the example shorter:

class MyStrategy(bt.SignalStrategy):
    params = (('period', 30),)

    def __init__(self):
        mysig = self.data.close > bt.indicators.SMA(period=self.p.period)
        self.signal_add(bt.signal.SIGNAL_LONG, mysig)

Which gets added to the mix as:


Any reader will notice that:

  • __init__ takes no parameters, named or not

  • There is no super call so the base class is not being directly asked to do its init

  • The definition of mysig references self.data which probably has to do with the YahooFinanceData instance which is added to cerebro

    Indeed it does!

There actually other attributes which are there and not seen in the example. For example:

  • self.datas: an array containing all data feeds which are added to cerebro
  • self.dataX: where X is a number which reflects the order in which the data was added to cerebro (data0 would be the data added above)
  • self.data: which points to self.data0. Just a ahortcut for convenience since most examples and strategies only target a single data

More can be found in the docs:

How are those attributes created?

In the 2nd article in this series it was seen that the class creation mechanims and instance creation mechanism were intercepted. The latter is used to do that.

  • cerebro receives the class via adstrategy

  • It will instantiate it when needed and add itself as an attribute

  • The new classmethod of the strategy is intercepted during the creation of the Strategy instance and examines which data feeds are available in cerebro

    And it does creates the array and aliases mentioned above

This mechanism is applied to many other objects in the backtrader ecosystem, in order to simplifly what the end users have to do. As such:

  • There is for example no need to constantly create function prototypes which contain an argument named datas and no need to assign it to self.datas

    Because it is done auto-magically in the background

Another example of this interception

Let’s define a winning indicator and add it to a winning strategy. We’ll repack the close over SMA idea:

class MyIndicator(bt.Indicator):
    params = (('period', 30),)
    lines = ('signal',)

    def __init__(self):
        self.lines.signal = self.data - bt.indicators.SMA

And now add it to a regular strategy:

class MyStrategy(bt.Strategy):
    params = (('period', 30),)

    def __init__(self):
        self.mysig = MyIndicator(period=self.p.period)

    def next(self):
        if self.mysig:
            pass  # do something like buy ...

From the code above there is obviously a calculation taking place in MyIndicator:

self.lines.signal = self.data - bt.indicators.SMA

But it seems to be done nowhere. As seen in the 1st article in this series, the operation generates an object, which is assigned to self.lines.signal and the following happens:

  • This object intercepts also its creation process
  • It scans the stack to understand the context in which is being created, in this case inside an instance of MyIndicators
  • And after its initialization is completed, it adds itself to the internal structures of MyIndicator
  • Later when MyIndicator is calculated, it will in turn calculate the operation which is inside the object referenced by self.lines.signal

Good, but who calculates MyIndicator

Exactly the same process is followed:

  • MyIndicator scans the stack during creation and finds the MyStrategy
  • And adds itself to the structures of MyStrategy
  • Right before next is called, MyIndicator is asked to recalculate itself, which in turns tells self.lines.signal to recalculate itself

The process can have multiple layers of indirection.

And the best things for the user:

  • No need to add calls like register_operation when something is created
  • No need to manually trigger calculations


The last article in the series shows another example of how class/instance creation interception is used to make the life of the end user easier by:

  • Adding objects from the ecosystem there where they are needed and creating aliases
  • Auto-registering classes and triggering calculations


Fork me on GitHub