## wpiea2026020-source-pdf

## Source details

**Canonical URL:** [wpiea2026020-source-pdf](https://www.imf.org/-/media/files/publications/wp/2026/english/wpiea2026020-source-pdf.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/wp/2026/english/wpiea2026020-source-pdf.pdf.md)
- [Structured JSON version](/-/media/files/publications/wp/2026/english/wpiea2026020-source-pdf.pdf.json)

---

### Executive Summary — Key findings
- Venezuela has lacked timely, reliable, publicly available key macroeconomic data by the Banco Central de Venezuela since Q1 2019, severely constraining accurate macroeconomic analysis and forecasting.
- The economy experienced a sharp and prolonged contraction, particularly between 2013 and 2020:
  - nominal GDP in 2013 shrank to a quarter of its 2012 size;
  - real GDP started to contract in 2014;
  - by 2020 real GDP had contracted to 75 percent relative to 2013.
- As of 2024:
  - real GDP had rebounded, growing on average at 4 percent;
  - the economy was still 70 percent below its 2013 level.
- Other salient socioeconomic facts:
  - Venezuela experienced a protracted hyperinflation episode between 2017-2021;
  - approximately 25 percent of the population has left the country (mostly to the region) as migrants and refugees;
  - there is a debt crisis with approximately US$ 140-190 billion of external debt in default.
- Modeling performance:
  - Introducing satellite data improved the Random Forest (RF) model by 13.85%;
  - the RF (non-linear) model outperformed the Dynamic Factor Model (DFM) by reducing the RMSE by 32.85%.
- Among regularization techniques considered, LASSO, Ridge and Elastic Net produced results inferior to Random Forest.

### Methodology and data
- Objective: close the data gap and nowcast quarterly real GDP for Venezuela up to 2024 by integrating traditional and non-traditional data with machine learning and econometric techniques.
- Target variable: real GDP (quarterly).
- Non-traditional data emphasized:
  - calibrated and harmonized nightlight datasets (time series spanning 2000–2024 for testing);
  - vegetation indices (NDVI, EVI);
  - NO₂ emissions;
  - other satellite-based indicators to proxy non-oil economic activity.
- Macroeconomic and other high-frequency inputs:
  - matching statistics for crude oil imports;
  - third-party sources for gas consumption;
  - revenue from VAT;
  - monetary aggregates;
  - self-curated macroeconomic variables from private sources.
- Modeling approaches evaluated:
  - Random Forest (RF) ensemble learning algorithm;
  - Dynamic Factor Model (DFM);
  - MIDAS-type models referenced in literature;
  - regularized linear methods: LASSO, Ridge, Elastic Net.
- Explainability tools used for RF:
  - Shapley values to assess predictor importance;
  - partial dependence plots to show marginal effects of features.

### Construction of a unified GDP series (Section 3.1 Real GDP)
- Official BCV real GDP data used up to 2019 Q1: seasonally adjusted real GDP from the Banco Central de Venezuela (BCV), covering 2006 Q1 to 2019 Q1.
- Private-source real GDP series used from 2019 Q2 onward: available from 1997 Q1 to 2023 Q4.
- Procedure to construct the unified series:
  - Seasonally adjust the private GDP series.
  - Compute year-on-year (Y/Y) and quarter-on-quarter (Q/Q) GDP growth rates for both sources to validate comparability.
  - Integrate BCV and private series into a composite Q/Q real GDP growth rate series that serves as the basis for the nowcasting model.
- Purpose and rationale:
  - Mitigate discontinuation of official GDP publication after 2019 Q1.
  - Address data discontinuities to enhance robustness of forecasting and nowcasting.
- Key numeric and temporal details (preserved exactly):
  - BCV coverage: 2006 Q1 to 2019 Q1.
  - Private series coverage: 1997 Q1 to 2023 Q4.
  - Unified series construction: private data from 2019 Q2 onward integrated with BCV up to 2019 Q1.
  - Composite Q/Q real GDP growth rate used as the basis for the nowcasting model.

### Results — Model performance and comparative findings (Section 5)
- Summary findings:
  - RF outperforms DFM in predictive accuracy, especially when satellite data are included, capturing nonlinearities and interactions among features.
  - DFM remains useful when integrating mixed-frequency economic data.
  - Validation used Time Series Split with a rolling-origin expanding-window cross-validation scheme; hyperparameter tuning was nested within this rolling scheme using a greedy grid search and monitoring of RMSE and MAE convergence across 5 folds.
  - Real-time publication lags are not available for Venezuela in the dataset; all predictors are used as observed.
  - Standard performance metric reported: Root Mean Squared Error (RMSE).
- 5.1 Random Forest Performance:
  - Training/testing sample split: 90:10 ratio.
  - Feature importance (without satellite data): crude oil production and gas consumption are the top features, followed by total capacity utilization (TCU) and monetary aggregates.
  - RF without satellite data:
    - RMSE: 6.51
    - Predicted one-step-ahead quarterly real GDP growth forecast (2024 Q1): 8.2 percent
  - Satellite variables included: nightlight, NDVI, EVI, and NO₂ emissions.
    - With satellite data, nightlight is among the top 4 features and the most important of the four satellite indicators.
    - Inclusion of satellite data moderates the dominance of crude oil production and gas consumption in the estimates; flaring can bias nightlight measures if not corrected.
  - RF with satellite data:
    - RMSE: 5.6
    - Predicted one-step-ahead quarterly real GDP growth forecast (2024 Q1): 8.5 percent
  - Interpretability diagnostics:
    - Shapley plot confirms nightlight as the strongest satellite predictor; EVI, NDVI, and NO₂ emissions show lesser explanatory power.
    - Partial Dependence Plot shows nightlight has the highest explanatory power among satellite-derived features and reveals nonlinearity, threshold effects, regime shifts, and structural breaks.
- 5.2 Dynamic Factor Model:
  - Training period: 2006 Q1 to 2021 Q4.
  - Testing period: 2022 Q1 to 2023 Q4.
  - DFM testing results:
    - RMSE: 8.34
    - One-step-ahead forecast of quarterly real GDP growth rate (testing dataset): 13.2 percent
  - Comparison:
    - DFM RMSE (8.34) is higher than both RF iterations (RMSE 6.51 without satellite data; RMSE 5.6 with satellite data) but lower than LASSO, Ridge, and Elastic Net Regression results reported in Annex I.
- Quantified improvements reported:
  - "The introduction of satellite data improved the RF model by 13,85% while the use of a non-linear model (in this case RF) outperformed the popular DFM model by reducing the RMSE by 32.85%."

### Advantages, limitations, and recommended extensions
- Advantages:
  - Satellite-based indicators provide indirect observation of economic activity where traditional data are unavailable.
  - Multi-source integration (satellite + macro) enables construction of a more comprehensive proxy for real GDP.
  - Techniques can highlight gaps or inconsistencies across series and improve monitoring in data-scarce environments.
- Limitations:
  - Satellite data were not originally designed for economic forecasting and require calibration (e.g., converting nightlight changes to economic activity).
  - Structural breaks and data disruptions necessitate methods robust to non-linearity and regime changes.
  - Experimental data sources and techniques cannot replace official statistics and require case-specific knowledge and regular ground truthing.
  - Nowcasting suffers similar update-frequency constraints as traditional sources when benchmark non-official private data are not frequently available.
- Suggested extensions and continued development:
  - Further calibration of satellite indicators to economic activity.
  - Integration of additional high-frequency non-traditional indicators.
  - Exploration of model explainability to enhance policy relevance and diagnostic use.
  - Integrate additional satellite indicators such as the Agricultural Stress Index (ASI) and Vegetation Health Index (VHI).
  - Improve feature engineering (e.g., include date information linked to business cycles).
  - Further hyperparameter tuning using different time-series cross-validation methods (fixed, rolling and expanding windows).

### Annex I — Comparison of Regression Techniques (A.4 Summary and Model Comparison)
- Comparative out-of-sample RMSE (lower is better):
  - Lasso Regression 5.74
  - Ridge Regression 6.22
  - Elastic Net 7.7
- Key insights:
  - LASSO Regression, with the lowest RMSE, indicates superior predictive capabilities while selecting the most relevant variables.
  - Ridge Regression performs the worst among the three in this application.
  - Elastic Net balances between Ridge and LASSO but yielded RMSE 7.7, higher than LASSO.
- Note: LASSO, Ridge and Elastic Net yielded inferior predictive performance relative to RF in the main results.

*International Monetary Fund — Executive Summary and selected sections (wpiea2026020-source-pdf)*

### Executive Summary ......................................................................................................

### Executive Summary

### Key findings
- Venezuela has lacked timely, reliable, publicly available key macroeconomic data by the Banco Central de Venezuela since Q1 2019, severely constraining accurate macroeconomic analysis and forecasting.
- The economy experienced a sharp and prolonged contraction, particularly between 2013 and 2020:
  - nominal GDP in 2013 shrank to a quarter of its 2012 size;
  - real GDP started to contract in 2014;
  - by 2020 real GDP had contracted to 75 percent relative to 2013.
- As of 2024:
  - real GDP had rebounded, growing on average at 4 percent;
  - the economy was still 70 percent below its 2013 level.
- Other salient socioeconomic facts:
  - Venezuela experienced a protracted hyperinflation episode between 2017-2021;
  - approximately 25 percent of the population has left the country (mostly to the region) as migrants and refugees;
  - there is a debt crisis with approximately US$ 140-190 billion of external debt in default.
- Modeling performance:
  - Introducing satellite data improved the Random Forest (RF) model by 13.85%;
  - the RF (non-linear) model outperformed the Dynamic Factor Model (DFM) by reducing the RMSE by 32.85%.
- Among regularization techniques considered, LASSO, Ridge and Elastic Net produced results inferior to Random Forest.

### Methodology and data
- Objective: close the data gap and nowcast quarterly real GDP for Venezuela up to 2024 by integrating traditional and non-traditional data with machine learning and econometric techniques.
- Target variable: real GDP (quarterly).
- Non-traditional data emphasized:
  - calibrated and harmonized nightlight datasets (time series spanning 2000–2024 for testing);
  - vegetation indices (NDVI, EVI);
  - NO₂ emissions;
  - other satellite-based indicators to proxy non-oil economic activity.
- Macroeconomic and other high-frequency inputs:
  - matching statistics for crude oil imports;
  - third-party sources for gas consumption;
  - revenue from VAT;
  - monetary aggregates;
  - self-curated macroeconomic variables from private sources.
- Modeling approaches evaluated:
  - Random Forest (RF) ensemble learning algorithm;
  - Dynamic Factor Model (DFM);
  - MIDAS-type models referenced in literature;
  - regularized linear methods: LASSO, Ridge, Elastic Net.
- Explainability tools used for RF:
  - Shapley values to assess predictor importance;
  - partial dependence plots to show marginal effects of features.

### Results and comparative performance
- RF with satellite data produced more accurate nowcasts than:
  - RF without satellite data (improvement of 13.85%);
  - Dynamic Factor Model (RF reduced RMSE by 32.85% relative to DFM).
- The presence of numerous structural breaks and the suspension of regular statistical publications in Venezuela limit the applicability of traditional techniques; the hybrid approach combining machine learning and satellite data better captures such disruptions.
- RF benefits from combining satellite data and high-frequency indicators; DFMs and other factor-based methods perform better when regular publication lags and recurring vignette releases are available, which is not the case for Venezuela in the examined period.
- LASSO, Ridge and Elastic Net were tested but yielded inferior predictive performance relative to RF.

### Advantages, limitations, and applicability
- Advantages:
  - Satellite-based indicators provide indirect observation of economic activity where traditional data are unavailable;
  - Multi-source integration (satellite + macro) enables construction of a more comprehensive proxy for real GDP;
  - Techniques used can highlight gaps or inconsistencies across series and improve monitoring in data-scarce environments.
- Limitations:
  - Satellite data were not originally designed for economic forecasting and require calibration (e.g., converting nightlight changes to economic activity);
  - Structural breaks and data disruptions necessitate methods robust to non-linearity and regime changes.
- Broader applicability:
  - The approaches may be useful for nowcasting and statistical improvement in other data-constrained economies, including fragile and conflict-affected states.

### Conclusion and recommendations
- Combining satellite data (especially nightlight and vegetation indices) with machine learning techniques, notably Random Forest, presents a viable method to nowcast real GDP in severely data-constrained contexts such as Venezuela.
- RF models that incorporate calibrated nightlight and other satellite indicators should be prioritized over linear regularization methods and, in many cases, over DFM when regular publication vignettes are absent.
- Continued development should focus on:
  - further calibration of satellite indicators to economic activity;
  - integration of additional high-frequency non-traditional indicators;
  - exploration of model explainability to enhance policy relevance and diagnostic use.

*International Monetary Fund — Executive Summary (wpiea2026020-source-pdf)*

### 3.1 Real GDP

### 3.1 Real GDP

### Construction of a unified GDP series
- Official BCV real GDP data used up to 2019 Q1: seasonally adjusted real GDP from the Banco Central de Venezuela (BCV), covering 2006 Q1 to 2019 Q1.
- Private-source real GDP series used from 2019 Q2 onward: available from 1997 Q1 to 2023 Q4.
- Procedure to construct the unified series:
  - Seasonally adjust the private GDP series.
  - Compute year-on-year (Y/Y) and quarter-on-quarter (Q/Q) GDP growth rates for both sources to validate comparability.
  - Integrate BCV and private series into a composite Q/Q real GDP growth rate series that serves as the basis for the nowcasting model.
- Purpose and rationale:
  - Mitigate discontinuation of official GDP publication after 2019 Q1.
  - Address data discontinuities to enhance robustness of forecasting and nowcasting.

### Validation and comparability
- Empirical comparisons:
  - Figure 1: Y/Y quarterly growth rate comparison shows a high degree of similarity between BCV and private series, supporting reliability of private estimates.
  - Figure 2: Q/Q growth rate comparison further reinforces viability of private data as a substitute.
  - Figure 3: Final composite Q/Q real GDP growth rate integrates both sources and is used for nowcasting.

### Role in nowcasting framework
- The composite Q/Q real GDP growth rate series is the core dependent variable for the nowcasting model.
- The unified GDP series enables combining traditional macroeconomic predictors and non-traditional (satellite) indicators to improve timeliness and accuracy in a data-scarce environment.

### Key numeric and temporal details (preserved exactly)
- BCV coverage: 2006 Q1 to 2019 Q1.
- Private series coverage: 1997 Q1 to 2023 Q4.
- Unified series construction: private data from 2019 Q2 onward integrated with BCV up to 2019 Q1.
- Composite Q/Q real GDP growth rate used as the basis for the nowcasting model.

*Source: INTERNATIONAL MONETARY FUND (wpiea2026020-source-pdf).*

### 5. Results

### 5. Results

### Summary findings
- Preliminary results suggest that RF outperforms DFM in terms of predictive accuracy, especially when satellite data are included, most likely due to its ability to capture nonlinearities and interactions among features.
- The DFM remains useful for nowcasting when integrating mixed-frequency economic data.
- Validation used Time Series Split with a rolling-origin expanding-window cross-validation scheme; hyperparameter tuning was nested within this rolling scheme using a greedy grid search and monitoring of RMSE and MAE convergence across 5 folds.
- Real-time publication lags are not available for Venezuela in the dataset; all predictors are used as observed, and no future information enters training or validation windows.
- Standard performance metrics are reported (Root Mean Squared Error, RMSE).

### 5.1 Random Forest Performance
- Training/testing sample split: 90:10 ratio.
- Feature importance (without satellite data): crude oil production and gas consumption are the top features, followed by total capacity utilization (TCU) and monetary aggregates.
- RF without satellite data:
  - RMSE: 6.51
  - Predicted one-step-ahead quarterly real GDP growth forecast (2024 Q1): 8.2 percent
- Satellite variables included: nightlight, NDVI, EVI, and NO₂ emissions.
  - With satellite data, nightlight is among the top 4 features and the most important of the four satellite indicators.
  - Inclusion of satellite data moderates the dominance of crude oil production and gas consumption in the estimates, though flaring can bias nightlight measures if not corrected.
- RF with satellite data:
  - RMSE: 5.6
  - Predicted one-step-ahead quarterly real GDP growth forecast (2024 Q1): 8.5 percent
- Interpretability diagnostics:
  - Shapley plot confirms nightlight as the strongest satellite predictor; EVI, NDVI, and NO₂ emissions show lesser explanatory power in this case.
  - Partial Dependence Plot shows nightlight has the highest explanatory power among satellite-derived features and reveals nonlinearity, threshold effects, regime shifts, and structural breaks—explaining improved RF performance relative to linear models.

### 5.2 Dynamic Factor Model
- Training period: 2006 Q1 to 2021 Q4.
- Testing period: 2022 Q1 to 2023 Q4.
- DFM testing results:
  - RMSE: 8.34
  - One-step-ahead forecast of quarterly real GDP growth rate (testing dataset): 13.2 percent
- Comparison:
  - DFM RMSE (8.34) is higher than both RF iterations (RMSE 6.51 without satellite data; RMSE 5.6 with satellite data) but lower than LASSO, Ridge, and Elastic Net Regression results reported in Annex I.

### Conclusion, implications, and recommended extensions
- Combining satellite data (vegetation cover, night/day light and nitrogen dioxide) with machine learning techniques (random forest and dynamic factor modelling) for Venezuela up to 2024 yields performance improvements over traditional methods and indicators and can help close data gaps in data-scarce environments.
- Quantified improvements reported in the paper:
  - "The introduction of satellite data improved the RF model by 13,85% while the use of a non-linear model (in this case RF) outperformed the popular DFM model by reducing the RMSE by 32.85%."
- Limitations and cautions:
  - Experimental data sources and techniques cannot replace official statistics and require case-specific knowledge and regular ground truthing.
  - Nowcasting suffers similar update-frequency constraints as traditional sources when benchmark non-official private data are not frequently available.
  - Changes in technology or accessibility of non-traditional data can impair model performance.
- Suggested extensions:
  - Integrate additional satellite indicators such as the Agricultural Stress Index (ASI) and Vegetation Health Index (VHI) to better proxy the agricultural sector beyond NDVI and EVI.
  - Improve feature engineering, e.g., include date information linked to business cycles to better capture trends, seasonal variation, and episodes such as hyperinflation.
  - Further hyperparameter tuning using different time-series cross-validation methods, comparing fixed, rolling and expanding windows, may enhance accuracy.

### Key statistics and exact figures
- Cross-validation folds: 5 folds
- Training/testing split used in RF experiments: 90:10 ratio
- RF without satellite data:
  - RMSE: 6.51
  - One-step-ahead forecast (2024 Q1): 8.2 percent
- RF with satellite data:
  - RMSE: 5.6
  - One-step-ahead forecast (2024 Q1): 8.5 percent
- DFM:
  - Training period: 2006 Q1 to 2021 Q4
  - Testing period: 2022 Q1 to 2023 Q4
  - RMSE (testing): 8.34
  - One-step-ahead forecast (testing dataset): 13.2 percent
- Reported proportional improvements:
  - Satellite data improved RF model by 13,85%
  - RF reduced RMSE relative to DFM by 32.85%

*Source: IMF — 5. Results (extract from provided PDF).*

### Annex I. Comparison of Regression Techniques

### Annex I. Comparison of Regression Techniques

### Overview
- Presents an overview of three regularized regression techniques: LASSO, Ridge, and Elastic Net Regression.
- Purpose: introduce penalty terms to Ordinary Least Squares (OLS) regression to address overfitting, multicollinearity, and high-dimensionality.
- Out-of-sample performance evaluation compares Root Mean Squared Error (RMSE) across models; lower RMSE indicates better predictive accuracy.

### A.1 LASSO Regression (L1 Regularization)
- Definition and purpose:
  - The Least Absolute Shrinkage and Selection Operator (LASSO) applies L1 regularization to a linear model.
  - Balances model simplicity and accuracy by encouraging sparsity (some regression coefficients forced to exactly zero), enabling feature selection.
- Optimization problem (as presented):
  - min_{β0,β1} ∑_{i=1}^{n} (y_i − β0 − X_i β)^2 + λ ∑_{j=1}^{p} |β_j|
- Notation and components:
  - y_i is the dependent variable.
  - X_i is the vector of independent variables.
  - β0 is the intercept.
  - β_j are the regression coefficients.
  - λ is the tuning parameter, controlling the degree of shrinkage.
- Properties and advantages:
  - As λ increases, more coefficients are shrunk to zero, effectively removing non-informative predictors.
  - Selects the most relevant variables by shrinking irrelevant coefficients to zero.
  - Penalizes large coefficients, reducing overfitting and improving model generalization.
  - Improves interpretability by eliminating redundant predictors.
  - Suitable for high-dimensional data, especially when the number of predictors exceeds the number of observations.
- Performance Evaluation (Out-of-Sample RMSE):
  - Figure referenced: "Figure 17. LASSO Regression: actual vs predicted real GDP growth rate" (Source: Calculations authors’ own).

### A.2 Ridge Regression
- Definition and purpose:
  - Ridge regression is an L2-regularized regression model to address overfitting and multicollinearity in high-dimensional datasets.
  - Unlike LASSO, Ridge shrinks all regression coefficients toward zero without eliminating variables entirely.
- Optimization problem (as presented):
  - min_{β0,β1} ∑_{i=1}^{n} (y_i − β0 − X_i β)^2 + λ ∑_{j=1}^{p} β_j^2
- Notation:
  - λ controls the extent of regularization.
- Properties and advantages:
  - Does not shrink coefficients to exactly zero; reduces their magnitude so all variables contribute.
  - Helps reduce overfitting and prevents excessive variance in model predictions.
  - Effectively handles multicollinearity when predictors are highly correlated.
  - Retains all predictors (no variable elimination).
- Performance Evaluation (Out-of-Sample RMSE):
  - Figure referenced: "Figure 18. Ridge Regression: actual vs predicted real GDP growth rate" (Source: Calculations authors’ own).

### A.3 Elastic Net Regression
- Definition and purpose:
  - Elastic Net Regression combines LASSO (L1) and Ridge (L2) regularization to balance variable selection and coefficient shrinkage.
  - Particularly effective with high-dimensional and highly correlated predictors.
- Optimization function (as presented):
  - min_{β0,β1} ∑_{i=1}^{n} (y_i − β0 − X_i β)^2 + λ1 ∑_{j=1}^{p} |β_j| + λ2 ∑_{j=1}^{p} β_j^2
- Notation:
  - λ1 controls L1 regularization (LASSO).
  - λ2 controls L2 regularization (Ridge).
- Flexibility between LASSO and Ridge:
  - If λ1 = 0, the model simplifies to Ridge Regression.
  - If λ2 = 0, the model simplifies to LASSO Regression.
- Properties and advantages:
  - Balances feature selection and shrinkage, retaining benefits of LASSO while reducing Ridge’s over-penalization.
  - Handles multicollinearity effectively; can perform better than LASSO when predictors are highly correlated.
  - Stable in high-dimensional data; useful when the number of predictors exceeds the number of observations.
- Performance Evaluation (Out-of-Sample RMSE):
  - Figure referenced: "Figure 19. Elastic Net Regression: actual vs predicted real GDP growth rate" (Source: Calculations authors’ own).

### A.4 Summary and Model Comparison
- Comparative findings (out-of-sample RMSE; lower is better):
  - Lasso Regression 5.74
  - Ridge Regression 6.22
  - Elastic Net 7.7
- Key insights:
  - LASSO Regression, with the lowest RMSE, indicates superior predictive capabilities while selecting the most relevant variables.
  - Ridge Regression performs the worst among the three, suggesting that coefficient shrinkage without feature selection leads to poorer generalization in this application.
  - Elastic Net strikes a balance between Ridge and LASSO, but its RMSE is slightly higher than LASSO, indicating that pure feature selection is more effective in this case.
- Performance evaluation figures and tables referenced are attributed to "Calculations authors’ own."

*Source: Annex I. Comparison of Regression Techniques (from the provided PDF content).*

---


_Source: https://www.imf.org/-/media/files/publications/wp/2026/english/wpiea2026020-source-pdf.pdf_