kernc
Iteration with Single-Step Progression to Improve Flexibility and Adaptability in Algorithmic Strategies
It is possible to split the run() function into two distinct functions, allowing for a single-step iteration on the ticks:
- initialize(): Saves the data and the strategy into a self variable.
- next(): Moves to the next tick in the iteration.
In which scenario do you want to use a next() function instead of a run()? Using a next() function allows passing any kwargs to the Strategy.next() function. This flexibility enables processing additional information outside the strategy class and injecting it as needed. However the run function is preserved in order to have the back-compatibility.
For example, one can utilize backtesting.py as an engine to train a reinforcement learning model, requiring an observation and an action at each step.
I've included a test/_iteration.py with a custom TestStrategy that works with both run and iteration, demonstrating how you can inject any parameters into the next function. Here an example how I use the single step iteration in order to training an Env with gym and stable-baseline3 library: https://gist.github.com/IperGiove/325832e30df44639ec78a618b84daf3c
Ok, I think I understand the issue. You can pass custom data to the strategy logic (df columns in addition to OHLC), but you can't pass custom objects (your RL optimizer object that tells your next action). In that case, you can:
df['model_obj'] = np.repeat(rl_model_obj, len(df))
and in next() use that (i.e. self.data.model_obj[-1]).
Would that work for you?
