yoshiki146 / japaneseeconforecast

Replication material for "Forecasting Japanese Macroeconomy Using High Dimensional Data"

Forecasting Japanese Macroeconomy Using High-Dimensional Data

Replication material for the JER paper, "Forecasting the Japanese macroeconomy using high-dimensional data"

Folders are organised as follows.

• data contains dataset, transformation codes, and their descriptions. (dat.rds is what you get in the end)

  • dat: Folders containing data before transformation. Filenames correspind to the sources or variable names (see datDetail.xlsx) . Only publicly available sources are uploaded (Datastream and NEEDS require subscription, hence only available for the transformed series). Abbreviations are: BOJ = Bank of Japan, METI = Ministery of Economy, Trade and Industry, MHLW = Ministry of Health, Labour and Welfare, MIC = Ministry of Internal Affairs and Communication, MLIT = Ministry of Land, Infrastructure, Transport and Tourism.
  • prep: Folders containing data preprocessing codes. Filenames correspond to the sources.
  • rds: Folders containing preprocessed data from prep. Stored in RDS format.
  • dat.r: puts the preprocessed data together for the analysis. Seasonal adjustments, unit root tests and data transformations are also done here.
  • dat.rds: Dataset to be used for the analysis.
  • preprocessing.r: executes all the codes at once
  • txt: Folders of some text files containing variables names to avoid writing long names in .r files
  • datDetail.xlsx: Data descriptions.

• models: Contain implementation codes of the corresponding model at a specific horizon and for a specific variable, defined in execution.r. Results can be found in results folder in rds format.

  • ar.r: Benchmark AR model. Lag orders are selected by BIC and kept track on ARlags. See result folder.
  • di.r: Diffusion index model with the number of factors selected using the information criterion by Bai and Ng (2002). DIfactor and DIlags record the number of factors and lag orders within a window, respectively. Note the number of factors are fixed irrespective of the target variables in this approach.
  • dicv.r: Diffusion indel model with the number of factors selected by h-step-ahead MSFE. DICVfactor and DICVlags record the number of factors and lag orders as in di.r. DICVr2 reports the R^2 of a regression of individual series onto estimated factor(s), which is to be used for the interpretation.
  • dilasso.r: Diffution index model with factors and lags selected by the lasso. DILASSOcoef is a list of 20 target variables within which contains sublist of three forecasting horizons. Each sublist contains 60 x 32 matrix (Number of window x number of factors & lags) with binary input; one means that the factor/lag is selected and zero not selected. DILASSOlambda reports the tuning parameter, lambda, selected by h-step-ahead MSE from the second subperiod. DILASSOnonzero reports the average number of selected parameters in a window and DILASSOr2 is R^2 of individual series onto selected factor(s) as in DICVr2.
  • lasso.r: Lasso implemented using glmnet package. LASSOcoefs reports if the variable is selected. Structured as DILASSOcoef with each sublist being 60 x 508 matrix, where 508 reflects 127 variables and 4 lags. LASSOlambda and LASSOnonzero is same as DILASSO.
  • enet.r: Elastic net with glmnet. ENETalpha contain the tuning parameter alpha to balance lasso-ridge weight. ENETlambda, ENETnonzero, and ENETcoefs are same as the lasso.
  • glasso.r: Group lasso with grplasso. gLASSOcoefs, gLASSOlambda, and gLASSOnonzero are kept track of.

• placeholders.r: Placeholders filled later in the implementation. Put in a different file so that execution.r remains neat.

• results: Contains results. MSFE contains MSFE in the absolute term (MSFEs.rds) and in the relative term (MSFE2.rds). See also the explanations for models above.

• txt: Text files containing the names of target variables to facilitate codes to be tidy.

• interpretations: Contains files used for interpretting the results. Used mainly for creating Fig 1, Fig 2, Table 5, and Table 6.