Estimating a Time-Varying VAR with Stochastic Volatility for Trading
Summary
The document introduces a Bayesian time-varying parameter vector autoregression with stochastic volatility (TVP-VAR-SV). Unlike a standard VAR with fixed coefficients and constant error variance, this model allows lag relationships, contemporaneous relationships, and volatility to evolve over time. The article describes the model’s variables, structural interpretation, prior construction from an initial data segment, and a mixture-of-normals approximation used to handle log squared shocks in the estimation process.
Estimation follows a Gibbs sampling approach associated with Primiceri and later work, combining state updates, Kalman filtering, and posterior draws. The article also describes using posterior draws to produce a next-day return forecast and says the model can inform a trading strategy, though the strategy details are largely absent from the supplied text. It offers background references and an output excerpt, but no complete performance evaluation, benchmark comparison, or evidence that forecasts are profitable. Results depend on model specification, priors, convergence, and the data used.
Key ideas
- A TVP-VAR-SV lets model coefficients and shock volatility change over time.
- The model can represent both lagged effects and contemporaneous relationships among series.
- Priors are estimated from an initial portion of the data before posterior inference on the remainder.
- A Bayesian Gibbs sampling procedure uses state estimation and mixture indicators to handle stochastic volatility.
- Forecasts derived from posterior draws do not by themselves establish trading profitability.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.