Research Econometrics

Euro GDP Interpolation

A quick experiment testing whether GDELT news sentiment can improve Chow-Lin temporal disaggregation of Euro Area GDP, extending Mönch & Uhlig (2005).

GitHub ↗
Python Econometrics Chow-Lin GDELT FRED Time Series
Questions

Two Questions, Clean Answers

1
Does news text add incremental information for monthly GDP interpolation?

No — text adds nothing beyond industrial production. Monthly R²: IP alone 0.965, IP + GDELT 0.965.

2
Can news text enable weekly GDP interpolation where IP cannot?

Partially — text alone produces a plausible path (R² = 0.94), reacts correctly to 5/6 major economic events and identifies turning points, but provides no marginal gain over IP.

Results

Model Comparison

Model Monthly R² Weekly R²
M0 — No auxiliary 0.500 0.500
M1 — IP only 0.965 0.991
M2 — GDELT only 0.593 0.940
M3 — IP + GDELT 0.965 0.991
M4 — IP + GDELT⊥ 0.965 —

Validation · Weekly GDELT GDP

Event Study
5/6 shocks correct direction
COVID Trough Detected
2020-05-11 (+14 days)
LOO-CV
No gain over mean
Alternative Methods
DFM, GP, Bayesian, XGBoost
Data

Sources

Quarterly Euro Area real GDP (FRED CLVMNACSCAB1GQEA19), monthly industrial production (FRED EA19PRINTO01IXOBSAM), and GDELT DOC 2.0 news coverage queried for "eurozone economy GDP recession inflation", cached on disk.

Pipeline

Implementation

1
Chow-Lin Estimator

From-scratch implementation with precision matrix and MLE; unit-tested against known outputs.

2
High-Frequency Extension

Daily/weekly Chow-Lin, residual-based variants, and a leading-indicator divergence measure.

3
Validation Suite

Leave-one-out cross-validation, Bry-Boschan turning-point dating, event study, variance decomposition.

Reference: Mönch, E. & Uhlig, H. (2005). "Towards a Monthly Business Cycle Chronology for the Euro Area." Deliverables: run.py pipeline, 13-page Beamer slides, 7-page article-format notes, and a full experiment log.