rstarBrookings2017

Replication files for Safety, Liquidity, and the Natural Rate of Interest by Marco del Negro, Domenico Giannone, Marc Giannoni, and Andrea Tambalotti, Brookings Papers on Economic Activity, Spring 2017: 235-294.

Updated r* estimates

The VAR Excel file contains the original as well as updated estimates of the trends in the real return for safe and liquid assets, the convenience yield, and its safety and liquidity components computed using the VAR model. These are the estimates shown in Figures 1 (black lines), 4, and 5 of the paper. The DSGE Excel file contains original as well as updated estimates of the 30-year, 10-year, and 5-year forward r* computed using the DSGE model. These are the estimates shown in Figures 1 (blue lines), and 12 of the paper. The figure below shows updated estimates of r* and the liquidity/safety component.

Required software

Julia v0.6.0 or above
DSGE.jl v0.4.1
MATLAB 16a

Download instructions

Download Julia from https://julialang.org/downloads/.
Open the Julia REPL and type:

a. Pkg.add("DSGE") to install DSGE.jl

b. Pkg.pin("DSGE", v"0.4.1") to use DSGE.jl v0.4.1

c. If, after running this replication code, you would like to use the most current version of DSGE.jl, type Pkg.free("DSGE") to un-pin the version

Installing this repository

Git users are welcome to fork this repository or clone it for local use. Non-Git users will probably find it easiest to download the zip file by clicking on the green Clone or download button on the right hand side of this screen, and then clicking "Download ZIP".

Directory structure

dsge/: Julia replication code for DSGE model
- spec1010_20_2016Q3_1221.jl: main script that generates all results
- input_data:
  - data/: transformed input data
  - user/: precomputed mode and hessian files
- output_data/m1010/ss20/:
  - estimate/:
    - raw/: raw estimation outputs: parameter draws from Metropolis-Hastings
    - tables/: LaTeX tables of parameter moments
  - forecast/:
    - raw/: full distribution, raw results (in model units) from all post-estimation products
    - work/: means and bands from raw results, transformed into final units, in binary format
    - tables/: LaTeX tables of post-estimation results (shock decomopositions, parameter histories)
tvar/: results for TVAR model
- MainModelX.m: main scripts that generate all the results
- DataCompleteLatest.xls: input data
- FiguresModelX/: output figures for each model specification
- Routines/: functions called to run estimation and produce figures
- output_data/: output files for all TVAR models
plot/: MATLAB code for plotting all results
- makeRstarPlots.m: Main driver script
- helperFunctions/: MATLAB functions called by makeRstarPlots
- Figures/: Output figure directory
- Tables/: Input table directory

How to run the DSGE code

The script spec1010_20_2016Q3_1221.jl generates all results for the DSGE model. As provided, the script will create a model object with the appropriate settings to estimate the model, compute smoothed histories of pseudoobservables, and produce shock decompositions ("pseudoobservables" is the term we use to describe linear transformations of states that we are interested in. In the Julia REPL with the DSGE package loaded, type ?pseudo_measurement for more information.

Three boolean variables at the top of the script indicate the operations the script will perform next. If run_estimation = true, the code will load the pre-computed mode and hessian files and run Metropolis-Hastings (be aware that this runs sequentially and could take more than 20 hours). To compute a mode and hessian from scratch, set reoptimize and calculate_hessian to true under "Settings for estimation" in the script. This will take much longer.

If run_modal_forecast = true, the code will load the provided parameter mode and compute smoothed pseudoobservable histories and shock decompositions. This runs quite quickly.

If run_full_forecast = true, the code will load all parameter draws from the estimation and compute smoothed pseudoobservables and shock decomositions for each draw. It will then compute means and bands across all draws. Since we compute results for 20,000 parameter draws, we parallelize this computation across 50 workers. Blocks of 500 draws each are sent to the workers, which compute and record results draw-by-draw. If you are not working with a cluster, set nworkers = 1 in spec1010_20_2016Q3_1221.jl (under "Parallelization"). If you have a cluster, set addprocsfcn appropriately for your machine (see ClusterManagers.jl).

How to run the TVAR code

The MainModelX.m scripts generate results for TVAR model (each script named according to the model number X specified in the paper). As provided, each script estimates the TVAR model for the given specification, outputs a .mat file of estimation results, produces figures for the distribution of trends, and prints the change in trends for the variables specified in the model.

If RunEstimation = 1, the code will run the estimation for the specified model. If RunEstimation = 0, the code will load the estimation results from a previous estimation. The first time the code is run, RunEstimation should be set to 1 to run the estimation and produce the necessary results.

We ran the estimation scripts in MATLAB R2016a.

How to run plots

See the README.md file in the plot directory. All results are plotted using MATLAB 16a.

erichuang49 / rstarbrookings2017 Goto Github PK