Prophetverse roadmap 2024-2025

[felipeangelimvieira]

This discussion will be a living collection of planned improvements in Prophetverse library for 2024 and 2025. I invite you to comment, ask, and suggest every aspect of the roadmap we are proposing.

Hierarchical model

• Add other reconciliation strategies, such as OLS and MinTrace. The MinTrace may be implemented by using a LKJ prior over the error covariance matrix, which could be optimized within the model.
• Add Gamma and Negative Binomial likelihoods
• Extend the support of some Effects to the hierarchical model.

Marketing Mix Modeling

• Add Adstock effect to prophetverse.effects
• Add a composition effect, so that the user may use Adstock before HillEffect as suggested in the literature and used other libraries.
• Add optimization module, so that users can optimize some input variable so that the target variable is maximize/minimized.
• Add visualization tools for Marketing Mix Modeling

Visualization

• Add visualization tools for MCMC samples (possibly an integration with arviz library) and for MAP optimization (loss curve)

Effects API

• Add functionality to easily run prior predictive check
• Add functionality to run effects individually with parameters obtained from an Inference Engine
• Add effects with Gaussian Process to model changing coefficients

Documentation

• Add example for MMM.
• Add more tutorials on how-to create custom effects.

General improvements

• Improve optimizer arguments when using MAP. Currently, it is not so user-friendly to choose different optimizers, since optimizer_kwargs are not the same for different ones. In addition, the interface with optax works but it's not good in current state
• Improve trend arguments in BaseProphetForecaster. Currently it is a ambiguous and can be improved.

[felipeangelimvieira]

@felipeffm FYI