In ooelrich/oscbvar: Stuff for Bayesian vector autoregressions

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>"
)

library(oscbvar)

This vignette goes through how to generate data for the paper: from (atomic) agent predictions, to the full data set including aggregation based methods.

Generate the agent predictions

Agent predictions are generated using the gen_atomic_preds function, which requires two inputs: a list of models for which to generate predictions for (a list of available models can be obtained using the gen_atomic_list() function) and a list of settings (the starting point, the minimum estimation size before prediction starts, whether to use a rolling estimation window or not, and which column contains the response variable). The function uses the macrodata set.

menu <- gen_atomic_list()
list_of_models <- menu[c(2, 6, 9, 10)]
atomic_settings <- list(5, 60, FALSE, 1)
pred_data <- gen_atomic_preds(list_of_models, atomic_settings)

This will return a data frame on the format

#?builhead(atomdat_2)

Generate aggregate predictions

Aggregate predictions come in two types. The first type is baseline predictions which are the two static pooling methods equal weights and gewisano (Geweke and Amisano's static otpimal linear pools). The second type is (dynamic) caliper pools.

The generation of aggregate predictions using the caliper methods goes through three steps. First we create a weight data frame using the function caliper_relevance, these weights are then combined with atomic predictions (historical data) to create what used to be call relevance-adjusted log scores. These relevance adjusted log scores are then used to create weights using the function gen_RAA, which then sends us to the propto_weighting function that calculates log predictive densities and predictive means. So propto_weighting calculates the final predictions, and as it does that it sort of calculates the weights as a by-product, adn thus we cannot really reach in there and get the weights out.

Taking a look at the weights over time

To get to the weights we need a data frame of atomic predictions, a data frame of smoothing variables, a time to start aggregating, a caliper width and a mvc. To get to the weights we simply need to first generate a weight matrix,then call RAL_calculator with that weight df and the atomic data frame, then manipulate the data frame we get back.

ooelrich/oscbvar documentation built on Sept. 8, 2021, 3:31 p.m.