replicate

#!/bin/bash

mkdir -p nlZs
mkdir -p results

# The pipeline of our software is summarized as below: 
# 1. Generate gp posteriors using main.m given forecasting
#   horizons, cross validation years and gp prior type.
#   For each horizon/cv year/type setting, there will be
#   100 files corresponding to 100 different random
#   hyperparameter setting using sobol sequence.
# 2. Compute averaged log likelihood on validation year using all year gp
#   posteriors given horizon, gp prior type and hyperparameter
#   index. Each DR model is trained on all years except
#   the year that is being cross-validated. There will be in 
#   total num of cv years * num of horizons * num of hyperparameter
#   settings such averaged log likelihood for each gp prior type.
#   NOTE: this step requires the majority of computation.
# 3. Determine the optimal hyperparameters for each 
#   horizon/type setting by comparing the log likelihood 
#   of the validation election year races, averaged over all
#   validation years.
# 4. Use the optimal hyperparameters to generate gp posteriors 
#   for test year (2018) and forecasting year (2020). Then generate trained
#   STAN objects using all cv year data and obtain Dirichlet regression 
#   posteriors, for each horizon.
# 5. Baseline: Bayesian random walk
# 

# define variables
horizons=(0 7 14 21 28 42 56)
cv_years=(1992 1994 1996 1998 2000 2002 2004 2006 2008 2010 2012 2014 2016)
gp_prior_types=("GP" "LM")


# 1. generate gp posteriors for Leave-one-year-out CV

# for TYPE in ${gp_prior_types[@]}; do
#     for horizon in ${horizons[@]}; do
#         # matlab code will iterate over 1992 to 2016 years
#         matlab -nodisplay -nodesktop -r "main('${TYPE}', 1, ${horizon}, 0); exit;"
#     done
# done

# 2. compute averaged log likelihood on validation year

# for TYPE in ${gp_prior_types[@]}; do
#     for horizon in ${horizons[@]}; do
#         for cv_year in ${cv_years[@]}; do
#             if [ ${TYPE}=='GP' ]; then
#                 num_hyps=100
#             else
#                 num_hyps=20
#             fi
#             for i in {1..${num_hyps}};
#                 Rscript --vanilla onejob.R $horizon $cv_year $TYPE $i
#             done
#         done
#     done
# done

# 3. determine the optimal hyperparameters 
# Rscript --vanilla loocv_nlZs.R

# 4. generate gp/DR posteriors for test year and forecasting year

test_years=(2018 2020)

# for TYPE in ${gp_prior_types[@]}; do
#     # matlab code will iterate over all horizons
#     matlab -nodisplay -nodesktop -r "main('${TYPE}', 2, 0, 0); exit;"
#     matlab -nodisplay -nodesktop -r "main('${TYPE}', 3, 0, 0); exit;"
# done

# for TYPE in ${gp_prior_types[@]}; do
#     for horizon in ${horizons[@]}; do
#         for test_year in ${test_years[@]}; do
#             Rscript --vanilla main.R  $test_year $TYPE $horizon
#         done
#     done
# done

# 5. baseline: bayesian random walk

test_years=(1992 1994 1996 1998 2000 2002 2004 2006 2008 2010 2012 2014 2016 2018)

for horizon in ${horizons[@]}; do
    for test_year in ${test_years[@]}; do
        Rscript --vanilla dynamicBayesian.R $horizon $test_year 
    done
done