bear_lite

A standalone, minimal wrapper around the BEAR toolbox for running a standard Bayesian VAR (BVAR) and computing the historical decomposition (HD).

This folder is fully self-contained and does not depend on any other file in the BEAR toolbox repository.

---

Files

bear_lite/
  bear_settings.m      – Settings constructor (returns a struct with defaults)
  bear_run.m           – Main entry point
  +bear/               – Internal package (copied verbatim from the BEAR toolbox)
    36 algorithm and utility files

What was removed

Everything not needed for standard BVAR + historical decomposition:

Removed	Reason
OLS VAR, Panel BVAR, SV, TVP, MF	Only standard BVAR (VARtype=2) is needed
FAVAR	Not needed
Forecasts / conditional forecasts	Not needed
FEVD	Not needed
Sign/IV restrictions (IRFt 4,5,6)	Not needed; Cholesky/triangular suffice
Mean-adjusted BVAR (prior=61)	Complex; omitted
All display/plotting functions	Replaced by returning a struct
Excel I/O (xlsread/xlswrite)	Replaced by direct matrix input
Grid search (hogs)	Not needed
Block exogeneity tables	Supported via settings.bex=0 default
gensample, gendates	Replaced by direct data matrix input

---

Setup

Add only bear_lite to your MATLAB path (do not add tbx/bear alongside it to avoid namespace conflicts with the +bear package):

addpath('/path/to/tbx/bear_lite')

---

Quick start

addpath('/path/to/tbx/bear_lite')

% --- Data ---
% Load your T-by-n data matrix (rows = time, columns = variables)
data = randn(120, 3);   % example: 120 observations, 3 variables

% --- Settings ---
s = bear_settings();    % start from defaults
s.prior = 11;           % Minnesota prior (see bear_settings.m for all options)
s.IRFt  = 2;            % Cholesky identification
s.lags  = 4;            % 4 lags
s.It    = 2000;         % Gibbs iterations
s.Bu    = 1000;         % burn-in

% --- Run ---
results = bear_run(s, data);

% --- Inspect HD ---
% hd_estimates is a (C+2)-by-n cell array.
% C = n_shocks(n) + initial_conditions(1) + constant(1) [+ exo if provided]
% Each cell is [lower; median; upper] (3 x T).

n = results.n;  % number of variables = 3

% Median contribution of shock 1 to variable 1 over the sample:
contrib_shock1_var1 = results.hd_estimates{1, 1}(2, :);   % row 2 = median

% Median contribution of the constant to variable 1:
contrib_const_var1  = results.hd_estimates{n+2, 1}(2, :);

With exogenous variables

data_endo = randn(120, 3);
data_exo  = randn(120, 2);   % 2 exogenous variables
results = bear_run(s, data_endo, data_exo);
% exogenous contribution is in hd_estimates{n+3, i}

---

HD cell array layout

Row index	Content
1 .. n	Contribution of structural shock j to var i
n+1	Initial conditions
n+2	Constant (when settings.const = 1)
n+3	Exogenous variables (when data_exo supplied)
C+1	Unexplained part (always zero for IRFt 1-3)
C+2	Portion explained by structural shocks

Each cell contains a 3-by-T matrix: [lower_bound; median; upper_bound]. Band width is controlled by settings.HDband (default 0.68).

---

Supported settings

Setting	Default	Description
prior	11	11–13=Minnesota, 21–22=Normal-Wishart, 31–32=Independent NW, 41=Normal-diffuse, 51=Dummy obs
IRFt	2	1=Unrestricted, 2=Cholesky, 3=Triangular
IRFperiods	20	IRF horizon (affects computation but not HD)
It	2000	Total Gibbs iterations
Bu	1000	Burn-in iterations
lags	4	Number of VAR lags
const	1	Include constant (1=yes, 0=no)
ar	0.8	AR prior coefficient (scalar or n-by-1 vector)
lambda1	0.1	Overall prior tightness
lambda2	0.5	Cross-variable weighting
lambda3	1	Lag decay
lambda4	100	Exogenous/constant prior tightness
lambda5	0.001	Block exogeneity shrinkage
lambda6	0.1	Sum-of-coefficients tightness (only if scoeff=1)
lambda7	0.001	Initial observation tightness (only if iobs=1)
lambda8	1	Long-run prior tightness (only if lrp=1)
bex	0	Block exogeneity (0=no)
scoeff	0	Sum-of-coefficients dummy (0=no)
iobs	0	Initial observation dummy (0=no)
lrp	0	Long-run prior (0=no)
HDband	0.68	Confidence band for HD
cband	0.68	Confidence band for coefficient estimates

---

Equivalence with the full BEAR toolbox

The algorithm files in +bear/ are exact copies of the originals. No core estimation code has been modified. Results obtained with bear_run are numerically identical to those produced by BEARmain with the same settings and data (given the same random seed).