README.md

In ooelrich/oscbvar: Stuff for Bayesian vector autoregressions

oscbvar

This is a package containing code and data for the paper local prediction pools.

Generating the predictions of the agents

The predictions of the agents, which I will sometimes refer to as "atomic" predictions (as opposed to aggregate predictions made with pooling methods) are generated using notebook function (this nomenclature is maybe outdated by now), one notebook exists for each model.

This is probably a stupid way of doing things, as I am basically creating a function (the notebook-generator) that contains all the code required to run a specific model, and then generates a bunch of predictions from that function. BASICALLY, what I should be doing instead is having a general "prediction generating function" that takes a specific model as input and then does a switch or something similar.

BUT, things being as they are, here is the logic:

• when we create predictions, we are going to want some settings to be the same for all the models. These "parameters" are stored in a list that we send to the notebook function called agc (atomic generation controllers) and include which observation to consider as the first observation (in terms of the variable t), whether to use a rolling estimation window or not, and which variable is the response variable.
• we also supply a data set. This should contain a column named t that contains "time index" since we are dealing with time series data in the application.
• The rest of the stuff that is supplied are settings for the specific model.

The chain of analysis

• Starts with raw data (only for the bikesharing example, for the macroexample we have precleaned data).
• From this raw data, a cleaning script is applied and the cleaned up version is stored as a data set in the package.
• From the cleaned data, predictions are made by a stable of models, and a data set of pooling variables is created.
• From the predictions and the pooling data, aggregate predictions are made.

This all implies a chain of debugging, where we want to be able to look at everything from the raw data to the predictions (analysing these also involves analysing the models, as we are only collecting some aspect of them when we save only the predictive density and predictive mean) and finally the aggregations. This would mean making a set of debugging tools that help us at each stage to figure out if what we are looking at is "reasonable".

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