Prophet re-integration in the new “plugin” based Tetsuo/Argonaut build went well.
ARIMA is a work in progress. This will ultimately be AUTO_ARIMA.
After learning about these a bit more, I had said it would support Prophet, CatBoost, ARIMA, LSTM, AUTO_ARIMA, SARIMAX.
I am revising this to:
- Prophet
- CatBoost
- AUTO_ARIMA
- LSTM
- XGBoost
- LightGBM
- 1D CNN
- A custom ensemble mode (likely XGBoost, 1D CNN, and a simple logistic regression as a baseline)
AUTO_ARIMA has been a huge disappointment so far.
The ensemble idea is admittedly ambitious. I don’t know if the Argonaut design can accommodate it efficiently. I know it can “do” the work, I don’t know that it can do it fast enough to be “consumable work”. Ensemble introduction might be the point where Tetsuo becomes a “true” distributed system to spread the load out.
Even if it does go that direction, Argonaut is designed to be split up into pieces like this, it would just need a rework of the IO layer, so, this was never out of scope.
Regarding AUTO_ARIMA, its general performance has been so bland that I don’t know that it’s even worth continuing pursuing. It’s fragile, it’s prone to hyperparameter tuning failures, and at this point it’s just included as a versatility talking point for the Tetsuo system in general. I might not even fully flesh out its model tuning capabilities and keep it at a “demonstrative only” status. I’m not getting the “this made progress” vibe from having introduced it as a supported model.