#124 State Space Models & Structural Time Series, with Jesse Grabowski

6 snips

Jan 22, 2025

Jesse Grabowski, PhD candidate at Paris 1 Pantheon-Sorbonne and principal data scientist at PyMC Labs, dives into the intricate world of state space models in time series analysis. He discusses the powerful adaptability of Bayesian methods in econometrics, emphasizing how they enhance forecasting accuracy. Grabowski highlights the balance between model complexity and simplicity, the significance of understanding trends, and the practical applications of innovations and latent states. Plus, he unwraps the role of the Kalman filter in managing dynamic data.

Ask episode

AI Snips

Chapters

Books

Transcript

Episode notes

ADVICE

Prior Setting Tip

Use pm.find_constraint_prior for easier prior setting.
Specify desired distribution properties instead of manually adjusting parameters, simplifying model development.

INSIGHT

State-Space Model Definition

State-space models are defined by a policy function, 'g', driving transitions between states.
Recursion is central, with current state 'xt' summarizing relevant history and 'g' mapping 'xt' to 'xt+1'.

ANECDOTE

Gaussian Random Walk

A Gaussian random walk, while simple, demonstrates state-space principles and has multiple representations.
It can be a cumulative sum of IID normals, a recursive process, or even a GP.

Get the Snipd Podcast app to discover more snips from this episode

Get the app

Proudly sponsored by PyMC Labs, the Bayesian Consultancy. Book a call, or get in touch!

Our theme music is « Good Bayesian », by Baba Brinkman (feat MC Lars and Mega Ran). Check out his awesome work!

Visit our Patreon page to unlock exclusive Bayesian swag ;)

Takeaways:

Bayesian statistics offers a robust framework for econometric modeling.
State space models provide a comprehensive way to understand time series data.
Gaussian random walks serve as a foundational model in time series analysis.
Innovations represent external shocks that can significantly impact forecasts.
Understanding the assumptions behind models is key to effective forecasting.
Complex models are not always better; simplicity can be powerful.
Forecasting requires careful consideration of potential disruptions. Understanding observed and hidden states is crucial in modeling.
Latent abilities can be modeled as Gaussian random walks.
State space models can be highly flexible and diverse.
Composability allows for the integration of different model components.
Trends in time series should reflect real-world dynamics.
Seasonality can be captured through Fourier bases.
AR components help model residuals in time series data.
Exogenous regression components can enhance state space models.
Causal analysis in time series often involves interventions and counterfactuals.
Time-varying regression allows for dynamic relationships between variables.
Kalman filters were originally developed for tracking rockets in space.
The Kalman filter iteratively updates beliefs based on new data.
Missing data can be treated as hidden states in the Kalman filter framework.
The Kalman filter is a practical application of Bayes' theorem in a sequential context.
Understanding the dynamics of systems is crucial for effective modeling.
The state space module in PyMC simplifies complex time series modeling tasks.

Chapters:

00:00 Introduction to Jesse Krabowski and Time Series Analysis

04:33 Jesse's Journey into Bayesian Statistics

10:51 Exploring State Space Models

18:28 Understanding State Space Models and Their Components