Dynamic factor models for Julia. Adapted from Bok et al. 2017, code.

using Pkg
Pkg.add(PackageSpec(url="https://github.com/dhopp1/DFMs.jl"))
using DFMs

1. To estimate a new model, two things are necessary:

• Dataframe to input to the model. Must include one column of type Dates.Date. N observations and R variables/columns (doesn't include date column).
• Block dataframe of dimension R x B, where B is number of blocks. Values are 1 or 0 for whether or not that variable is to be included in the block.

2. run output_dfm = estimate_dfm(df; blocks=blocks_df, p=1, max_iter=5000, threshold=1e-5), where:

• p is number of lags for the AR dimension of the model
• max_iter is the max number of times to run the EM
• threshold is the threshold for convergence in the EM

3. save the estimated model parameters with export_dfm(output_dfm=output_dfm, out_path="out_folder"). The model can now be used for predictions on new datasets where columns remain the same as the dataset it was estimated on.

1. either estimate a new model following the instructions above, or load a previously estimated one saved via export_dfm as follows:

• output_dfm = import_dfm(path="out_folder")

2. obtain predictions which outputs a dataframe with predicted values for all series:

• predictions = predict_dfm(new_df; output_dfm=output_dfm, months_ahead=3, lag=0), where:
  • new_df is the new dataset with same columns as that the model was estimated on
  • output_dfm is the estimate_dfm output either newly estimated or loaded
  • months_ahead is the number of months forward to forecast
  • lag is the number of lags for the kalman filter (default 0)

1. news = gen_news(old_y=old_df, new_y=new_df, output_dfm=output_dfm, target_variable=:target_col, target_period=Dates.Date(2020,1,1)), where:

• old_y/new_y are the old and new datasets (i.e. same data, new_y has more recent values filled in)
• output_dfm is the output of a previously estimated model
• target_variable is which column you would like to forecast
• target_period is which date you would like to forecast for

2. Outputs a dictionary with :y_old, :y_hat, :news_table, run ?gen_news for more info.