Trying to extend IndicatorFactory instance

[Sanjan611]

I'm using the IndicatorFactory's from_pandas_ta method to get an instance of the TSI indicator (https://github.com/twopirllc/pandas-ta/blob/main/pandas_ta/momentum/tsi.py).

And so far I've got

TSI = vbt.IndicatorFactory.from_pandas_ta('tsi')

# and for some parameters - 
tsi = TSI.run(df['Close'],
            fast = fast_windows,
            slow = slow_windows,
            drift = drifts,
            offset = offsets,
            param_product = True)

Accounting for the the implementation of TSI in pandas_ta, the only output data I get is tsi.tsi. However, I want to generate signals based on the crossover of a signal (EMA of TSI) and that of the TSI. I've been racking my brains on how to create something like tsi.signals without too much rewriting. That way, I assume I would be able to do something like entries = tsi.tsi_above(tsi.signals, crossover = True).

@polakowo how can I add a new attribute to my tsi IndicatorFactory instance (if it is possible)?

I've also realised that I can't modify tsi.tsi - is it meant to like a frozen attribute?

[polakowo]

Generally, you shouldn't overwrite instance variables in vectorbt because this may cause side effects (e.g. in caching). That's why most attributes are read-only properties.

You can extend any indicator by subclassing it. But as soon as you need to integrate another parameter, you would need to create a new indicator.

[Sanjan611]

I see. So for my situation, if I wanted to incorporate something like tsi.signals into the indicator, I would need to create a new IndicatorFactory instance that does what I want it to do, and also return tsi.signals with tsi.tsi, right?

[polakowo]

As said, if your signals require a new parameter, then yes. Otherwise, simple subclassing should work.

class CustomTSI(TSI):
    def signals(self, ...): ...

If you need to create a new indicator upon any other pandas-ta indicator, look at how vectorbt does it -

vectorbt/vectorbt/indicators/factory.py

Line 3429 in fa021bb

@classmethod

[Sanjan611]

I've managed to something set up all fine. Would love your thoughts on this implementation! I figure I may not have done things in the most efficient way (still learning my way through). As I haven't yet gotten into the weeds of how something like tsi_above() is implemented, I've just written another method for entries and exits for now. I've also tried to use vectorbt's functions (that work on arrays only) to make the methods fast.

class CustomTSI(TSI):

    def get_signals(self, windows):
        """Method to get the MA of self.tsi 
        windows: List of window sizes for rolling mean"""

        # Convert DataFrame to array first
        tsi_arr = self.tsi.to_numpy()

        # Create a list of MA-processed arrays
        tsi_list = []
        for w in windows:
            tsi_list.append(vbt.nb.rolling_mean_nb(tsi_arr, w))

        # convert the list of arrays to a list of DataFrames
        dfs = [pd.DataFrame(tl, index = self.tsi.index, columns = self.tsi.columns) for tl in tsi_list]

        # build two new attributes that have `signal` as part of their columns MultiIndex 
        self.signals = pd.concat(dfs, axis = 1, keys = windows, names = ["signal"])
        self.tsix = pd.concat([self.tsi]*len(windows), axis = 1, keys = windows, names = ["signal"])

    def get_enex(self, crossover = False):
        
        self.entries = self.tsix > self.signals 
        self.exits = self.tsix < self.signals

        if crossover == True:
            self.entries = self.entries.vbt.signals.nst(1)
            self.exits = self.exits.vbt.signals.nst(1)
        
        return self.entries, self.exits

[polakowo]

@Sanjan611 there are several improvements you can make.

1. most generic functions such as rolling_mean_nb are available as accessor functions (see here). You can do self.tsi.vbt.rolling_mean(w) and it will automatically convert to numpy, call rolling_mean_nb, and convert back to pandas. The other advantage is that it doesn't care if input is series or dataframe (in your example, you call rolling_mean_nb which works on dataframes only, but self.tsi can be a Series)
2. for self.signals, there is this accessor function that lets you concat conveniently
3. for self.tsix, there is this function that lets you tile self.tsi a number of times
4. ideally if a method is prefixed with get_, it should return outputs instead of writing them to the object
5. generally, I prefer objects to be atomic and read-only. This is because most objects in vbt have methods copy, save, etc. so you can save them to disk. Those methods do not keep track of attributes (such as self.entries, etc.) that you added/changed! It's always best to keep your code pure (that is, without side effects). Looks like window parameter is a good reason to create a new indicator on top of TSI instead of subclassing it.

[Sanjan611]

Thanks for these! I'm just now realising the numerous accessor functions available and they all look really convenient to use.