Prophetverse

Prophetverse is an open-source Python library for Bayesian time series forecasting and Marketing Mix Modeling (MMM). It builds on the theoretical foundations of the original Prophet model and extends them into a more flexible and expressive framework with custom priors, non-linear effects, and multiple likelihoods.

Built on top of sktime and numpyro, Prophetverse focuses on interpretability, customizability, and modern Bayesian modeling practices.

Purpose

Many forecasting tools are either too rigid or difficult to adapt to domain-specific knowledge. Prophetverse addresses this gap by enabling:

• Explicit Bayesian modeling with interpretable parameters

• Integration of expert knowledge through custom priors

• Flexible model structures for complex real-world time series

A particular emphasis is placed on Marketing Mix Modeling, where understanding the causal contribution of different marketing channels is critical.

Core Features

• Flexible Bayesian time series models

• Native compatibility with the sktime ecosystem

• Support for multiple likelihood functions

• Customizable trends, seasonalities, and priors

• Non-linear effects for exogenous variables

• Strong focus on model transparency and interpretability

Modeling and Forecasting

Prophet-Compatible Interface

Prophetverse provides an interface compatible with sktime, making it easy to adopt for users familiar with traditional forecasting workflows while offering significantly more modeling flexibility under the hood.

Advanced Trend Modeling

Compared to the original Prophet model, Prophetverse supports:

• Logistic trends with capacity modeled as a random variable

• Fully customizable trend functions

• More stable changepoint handling over long time horizons

Probabilistic Modeling

Supported Likelihoods

To better model real-world business data, Prophetverse supports several likelihood functions:

• Gaussian

• Gamma

• Negative Binomial

• Beta

This enables robust forecasting for positive-only data, count data, and other non-Gaussian distributions.

Custom Priors

Users can define custom prior distributions to:

• Encode domain knowledge directly into the model

• Enforce constraints such as positive coefficients

• Improve model stability and interpretability

Exogenous Variables and Marketing Mix Modeling

Non-Linear Exogenous Effects

Prophetverse allows fully customizable, non-linear effects for exogenous variables, enabling sophisticated modeling of relationships between external drivers and the target time series.

Shared Coefficients and Hierarchies

Hierarchical and multivariate models with shared coefficients allow global information to improve individual forecasts, which is especially valuable in Marketing Mix Modeling scenarios.

Seasonality and Scaling

Flexible Seasonality

Seasonal patterns are modeled using Fourier terms passed as exogenous variables. This approach enables:

• Easy creation of custom seasonalities

• Multiple overlapping seasonal components

• Domain-specific periodic patterns without hardcoded assumptions

Scaling Strategy

• Target variables are scaled internally for numerical stability

• Exogenous variables remain under user control

• Seamless integration with sktime transformers

Multivariate Models

Prophetverse supports hierarchical multivariate models with:

• Multivariate normal likelihoods

• LKJ priors for correlation structures

• Bottom-up forecasting across related time series

Use Cases

• Interpretable time series forecasting

• Marketing Mix Modeling

• Sales and demand forecasting

• Scenario analysis and simulation

• Research and applied Bayesian modeling

Open Source

Prophetverse is released as open-source software and actively developed within the Python time series ecosystem. It leverages modern probabilistic programming via numpyro and integrates tightly with sktime.

GC.OS supports Prophetverse as an open-source project enabling transparent, flexible, and scientifically grounded forecasting and marketing analytics.