Discussion of Suero (2025)

BSP Research Huddle

2025-05-21

Thanks

  1. Harold for the invitation

  2. Joshua for the chance to read some work on forecasting

    • led to thinking about some connections to work-in-progress about combining data from probability and non-probability samples
    • one fo the figures in the paper motivated me to think about forecast combinations with large and non-interpretable model universes

My dips into forecasting – supervision

  1. 2 bachelor’s theses on fixed-\(T\) panel data forecasting

    • working out theoretical details of a base case
    • application

  2. 1 master’s thesis on fixed-\(T\) panel data forecasting

    • forecasting using correlated random effects

  3. 1 master’s thesis on forecasting aggregated NPL ratios in the Philippines

    • reverse MIDAS application

Summary

  1. Goal: Forecast total external debt in the Philippines

  2. Purpose: Attempt to provide a standardized workflow

  3. Presumed audience:

    • Central bank staff: For easy and fast deployment
    • The proactive policy-maker: Lessen reliance on one misspecified model

  4. (New-ish?) Time-varying optimal weighting of forecasts

  5. (New-ish?) Using the lasso for recovering weights

External debt?

  1. Uncommon to hear forecasts of the level of external debt

    • External debt as a percent of GDP?
    • Internal reports to BSP seems to contain decomposition of external debt
    • What are the reactions of economic agents to external debt projections?

Missing details

  1. The “REGRESS” step of the workflow is unclear, especially Table 2.

    • Is the data a set of forecasts produced by each model?
    • What are the regressors?
    • Is this the Aiolfi and Timmermann (2006) approach?
    • Where is the equally-weighted forecast benchmark?

Sanity check

  1. Standardize data prior to applying OLS for Table 2 in order to compare with the very large coefficients found in Figures 7 and 8.

    • Table 2 suggests that random walk with drift and support vector regressions “survive”.
    • Negative weights?
    • The “broken record”: How was \(\lambda\) chosen?
    • Results not surprising that random walk with drift “survives”.

Nitpicking

  1. “The quality of the forecasted data meets internationally recognized standards” – are forecasts of external debt required to comply with external debt reporting standards?

  2. MAPE was chosen: preferences of the forecaster? preferences of the forecast user?

  3. Confused about the language: selection or combination?

  4. Details of the methodology: important for unsupervised usage by central bank staffers

  5. Chosen colors for Figures 7 and 8

Recommendations

  1. Monte Carlo experiment to evaluate proposed workflow

  2. Literature review

    • Missing discussion of experience about debt projections
    • Forecast combination literature was ignored

  3. Clarity regarding

    • Data inputs
    • Workflow
    • Forecast intervals

Thanks for listening!