From model 8i1 the following semantic versioning of models is used. Major (number): No model back-compability, i.e. (some) previous model cannot be used. Minor (letter): New models/feature/inputs are added, but previous models can be run and should have the sam lpd. Patch (number): No changes in input or ourput or new flexibilities. Only code performance.
A simple state space model with cauchy priors
The state-space expanded with handling also multiple time points in polls and model that explicitly using time_weights of each poll.
The state-space model is expanded with scaling with time_scale_length.
Changed model 3 to use student t and nu to model the movement over time.
State-space with partly known information on the underlying state.
Extend the model to multiple parties, i.e. a multivariate model.
Handle missing values in data (lex Spain) and different length of parties (lex Germany).
As model 6b, but with the following additions: In this model we add industry bias by introducing a linear effect through kappa, a (T_known + 1) by P parameters. There exist one kappa parameter per party and election and the prior is centered around 0 (no industry bias). We learn a hiearchical prior parameter for all parties called sigma_kappa that has a truncated (positive) normal prior with sd = sigma_kappa_hyper. Hence sigma_kappa_hyper is our global parameter for how large we believe sigma_kappa can be for all parties and all elections.
As model8a, but with an intercept g and increasing b instead of an increasing g starting from zero.
As model8a, but with the last kappa assumed to be 0. This is done to test the model performance.
As model8a, but g is an real valued parameter instead of a real parameter. The interpretation of g is now the number of years since the last election. This is the best working model including industry bias.
As model8a3, but there is one sigma_kappa per party instead of one for all parties.
As model 6b, but with house effects on the mean, by house, party and slower moving time period s. The house biases has a dynamic changing over the time points s.
As model 8b, but where there is a sum-to-zero constraint over all houses.
As model 6b, but with house/design effects on the poll variances, by house, party and slower moving time period s. The design effects is dynamic changing over the time points s.
As model 6b, but with house/design effects on the poll variances, by house and slower moving time period s. The designe effect is slowly changing over s, but there is only one parameter per house.