Is loo a good package for Autoregressive model?

[Yufei_WU]

Hi :)

I have a model looks like this:

It is a variation of autoregressive model, and the sum up term take everything that happened before take into account, with a γ (a forgetting factor between 0~1, namely give the more recent events higher weights).

And I get high pareto k values (Most of them are bad to very bad) when fitting it with loo, but the posterior prediction checking looks good, and I don’t know where to start troubleshooting with it.

I was wondering does this indicate that loo is not a good package for my model? Or that my model is very badly misspecificed?

Thank you very much!

[maxbiostat]

@avehtari

[avehtari]

@Yufei_WU

And I get high pareto k values (Most of them are bad to very bad) when fitting it with loo, but the posterior prediction checking looks good,

If the model is overfitted, and PSIS fails, then you get high Pareto k values, and the posterior predictive checking looks good as the model is overfitting.

To say something more, can you post the loo output diagnostic and your model code?