Spike: Consider Using Darts for Managing Prophet and Other Forecasting Models #29
Description
Investigate the feasibility and benefits of using Darts to manage forecasting model providers.
Darts is a framework for time series forecasting that supports multiple models (including Prophet, XGBoost, LightGBM, etc.) with a consistent API. It also includes built-in support for regressors and common time series preprocessing, which may reduce the need for custom feature engineering.
Goals
- Evaluate how Darts can simplify model management
- Compare current Prophet implementation with a Darts-based equivalent
- Explore support for future model types via Darts (e.g. XGBoost, RNNs)
Notes on Pre-processing
Before Darts can be used, data pre-processing needs to be properly set up:
- Darts expects all time series data (target and covariates) to be resampled to a fixed interval
- Missing data must be handled explicitly during pre-processing
- Current Prophet implementation handles much of this internally, but Darts shifts this responsibility to the service
While this isn't a large amount of work, it's important to account for during migration.
Benefits
- Easy addition of new forecasting ML/stat models
- Native support for time series preprocessing and regressors
- Built-in backtesting and evaluation/validation utilities
- Potential reduction in boilerplate and custom logic