Monthly Task RollingGen

[pop0121]

🌟 Feature Description

Is it possible to provide a way to rolling train the model (and consequently, the backtest) at an irregular interval (e.g., weekly or monthly).

Sorry to post it if there is already a way to do it in qlib.

Motivation

• In many scenarios, investors may rebalance the portofolio weekly or monthly.
• A natural way to backtest the model and strategy is to simulate this process, and fine-tune the pre-trained model just before the re-balancing date, based on the latest obervations.
• However, the current version of qlib seems to only support regular interval rolling training (e.g., 20 days), which is not perfectly aligned with the trading calendar.

Alternatives

Additional Notes

[you-n-g]

The default implementation is a fixed_step version.
But it is free for users to customize its task generator for flexible interval
Users can implement a rolling generator by overriding this function.
And then it will be used in

qlib/qlib/workflow/online/strategy.py

Line 189 in 144e1e2

res.extend(self.rg.gen_following_tasks(task, calendar_latest))

[pop0121]

Thanks for you suggestiong.

I am trying to figure out if it is possible to do this by modifying the executor. I always try to avoid using task generator due to limited hardware resources, though it seems to be a good practice to parallelize different task settings.

I have managed to run a rolling-test backtest by following the example on "rolling dataset" instead of "rolling task". But, as another post says, it seems that the backtest part of qlib is not as well documented as the other parts.