Hi PyMC Community and Mentors,
My name is Mithilesh, and I am preparing a proposal for Google Summer of Code 2026 to work on expanding the pymc.timeseries module by implementing a modular API for Bayesian Structural Time Series (BSTS) and State Space Models.
While PyMC handles basic autoregressive distributions well, building composable State Space Models (like local level, seasonality, and stochastic volatility) currently requires writing custom, low-level PyTensor scan operations. My project aims to abstract this, creating a high-level API similar to R’s BSTS to make probabilistic financial forecasting more accessible.
To familiarize myself with the specific bottlenecks of the current architecture, I spent some time building a Proof of Concept: a Bayesian Stochastic Volatility model applied to EUR/USD exchange rates.
- Repository & Notebook: GitHub - Mithil-7/GSoC-2026-PyMC-Preparation: GSOC-2026 · GitHub
Building this manually highlighted exactly why this GSoC project is needed—dealing with the geometry of the posterior and the resulting Effective Sample Size (ESS) drops using standard distributions really underscored the need for an optimized, out-of-the-box structural time-series API!
I have drafted my full GSoC proposal outlining a 350-hour timeline, technical mitigations for PyTensor graph compilation, and my background in deep learning and quantitative finance.
I would be incredibly grateful for any feedback from the mentors or community members. Specifically, I would love to know if my timeline for Phase 3 (Multivariate Expansion) is realistic, or if I should narrow the scope strictly to univariate composability.
Thank you for your time and for maintaining such an incredible library!
Best regards, Mithilesh GitHub: Mithil-7 (Mithilesh A ) · GitHub