search

nicholasjclark / mvgam

{mvgam} R 📦 to fit Dynamic Bayesian Generalized Additive Models for time series analysis and forecasting
https://nicholasjclark.github.io/mvgam/
Other
91 stars 11 forks source link

Allow hierarchical or no pooling of trend components #35

Open nicholasjclark opened 7 months ago

nicholasjclark commented 7 months ago

Right now autoregressive and trend variance parameters are 'hierarchical', but the hyperparameters are are fixed (i.e. ar1 ~ normal(0, 0.5)). It would be useful to allow options to learn these hierarchically, i.e.

ar1 ~ normal(ar1mu, ar1sigma);
ar1mu ~ normal(0.5, 0.1);
ar1sigma ~ exponential(5);

This is probably more relevant for variance parameters as different series may have wildly different dynamics

nicholasjclark commented 2 months ago

If this goes ahead it'll undoubtedly need the noncentred parameterisation