## Nowcasting GCC GDP: A Machine Learning Solution for Enhanced Non-Oil GDP Prediction

## Source details

**Canonical URL:** [Nowcasting GCC GDP: A Machine Learning Solution for Enhanced Non-Oil GDP Prediction](https://www.imf.org/-/media/files/publications/wp/2025/english/wpiea2025268-source-pdf.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/wp/2025/english/wpiea2025268-source-pdf.pdf.md)
- [Structured JSON version](/-/media/files/publications/wp/2025/english/wpiea2025268-source-pdf.pdf.json)

---

### Overview
- Introduces a real-time nowcasting framework for estimating quarterly non-oil GDP growth in the GCC.
- Provides high-frequency, sector-specific estimates to give policymakers timely insight into GCC non-oil economic activities.
- Extends nowcasting literature for the MENA region by moving beyond single-model approaches and incorporating a richer set of high-frequency, cross-border indicators.
- First systematic application of a unified nowcasting framework across the GCC with country-specific models for non-oil GDP.

### Methodology and Framework
- Flexible, machine learning-based, model-agnostic setup encompassing 22 candidate models (see Annex II for model list).
- Aggregates estimates across factor-based regressions, bridge equations, and machine learning techniques.
- Iterative design: incorporates and reweights incoming information as new data are released; accommodates frequent data revisions and evolving statistical capacity in the GCC.
- Two key innovations:
  - Advanced data collection and integration tailored to the GCC combining traditional metrics with real-time financial transactions and select unconventional data.
  - Novel use of Shapley value decompositions to quantify marginal contribution of each predictor and iteratively refine indicator selection.

### Data and High-frequency Indicators
- Indicator categories: consumer prices and subcomponents, PMIs and industrial production, trade data, point-of-sale transactions, credit, deposit rates, stock market indicators, oil price benchmarks, interest rate spreads, financial flows from key partners (United States, China, Saudi Arabia, UAE).
- Unconventional sources considered (limited use): nighttime lights, land cover, nitrogen dioxide concentration, PortWatch trade, google search trends, satellite imagery, social media contents.
- Variable selection: from millions of series down to ~1,000 country-relevant series per economy, then top ~50 indicators by historical correlation, with final sets of approximately 10–15 indicators per economy.

### Timeliness and Data Publication Lags (Selected Statistics)
- Target publication lag for indicators: no more than ±45 days.
- Quarterly GDP publication lags (from the reference quarter) and available time series (as of June 2025):
  - Bahrain: 2008Q1 – 2024Q4 +90 days
  - Kuwait: 2010Q1 – 2024Q4 +90 days
  - Oman: 2010Q1 – 2024Q4 +105 days
  - Qatar: 2011Q1 – 2024Q4 +90 days
  - Saudi Arabia: 2011Q1 – 2025Q1 +68 days
    - Note: GASTAT publishes flash estimates with a 30-day lag and official data with a 68-day lag. Data vintage as of end-April 2025 was used for Saudi Arabia in comparisons.
  - United Arab Emirates: 2012Q1 – 2024Q3 +200 days [varies]

### Variable Consideration, Selection Criteria, and Transformations
- Core selection criteria:
  1. Predictive Capability: typical threshold correlation coefficient greater than ±0.7 (threshold adjusted by country).
  2. Historical Coverage: indicators expected to have at least 10 years of historical data.
- Transformations and frequency handling:
  - Daily and weekly indicators converted to monthly via aggregation or end-of-period selection.
  - Variables transformed into year-on-year growth rates for standardization.
- Timeliness trade-offs:
  - Prioritize indicators with publication lags ≤ ±45 days where possible.
  - Indicator basket typically locked for subsequent iterations; series with two or more consecutive quarters of negligible Shapley contribution are revisited.

### Nowcasting Model Evaluation and Selection
- Three-stage process:
  1. Training: models trained on historical dataset using selected 10–15 high-frequency predictors and at least 10 years of data; target = non-oil real GDP.
  2. Holdout test: predictive performance assessed on holdout test set using Root Mean Squared Error (RMSE); preliminary filter applied.
  3. Cross-validation and final selection: rolling window approach typically reserving final 15 percent as holdout; best-performing model by out-of-sample RMSE re-estimated on full dataset for final nowcasts.
- Algorithms emphasized:
  - Support Vector Machines (SVMs), Random Forests, Stochastic Gradient Boosting Trees (GBM), Elastic Net, Principal Component Regression (PCR), and others.
- Model-agnostic approach: select most appropriate algorithm per country and data environment to ensure accuracy, adaptability, generalizability, and interpretability.

### Shapley Value Decompositions and Interpretability
- Shapley values quantify each variable’s marginal contribution to predictive accuracy and are used diagnostically to:
  - Flag low-contribution indicators for exclusion.
  - Inform variable selection across dozens of model structures.
- Practical outcomes:
  - Final tailored sets: approximately 10–15 indicators per economy, each demonstrating consistent contribution.
- Interpretation note:
  - Shapley values indicate shared predictive influence among correlated variables and are not causal effects.

### Key Results and Contributions
- Produces high-frequency non-oil GDP nowcasts that enhance granularity, responsiveness, and transparency of short-term forecasts.
- Identifies key high-frequency indicators strongly correlated with non-oil activities in the GCC.
- Complements institutional forecasts; for Saudi Arabia, the model slightly outperformed official non-oil flash estimates in sample average of forecast errors.

### Nowcasting Statistics — Top 3 Performing Models and Out-of-Sample RMSE (Q1 2025; forecasting non-oil GDP growth for Q2 2024)
- Bahrain (as of June 12, 2025)
  - 1. Stochastic Gradient Boosting Trees (oneSE) — 0.72
  - 2. Stochastic Gradient Boosting Trees — 1.07
  - 3. Random Forest (oneSE) with Variable Selection — 1.35
- Kuwait (as of June 12, 2025)
  - 1. Support Vector Machine (RBF) — 2.50
  - 2. Support Vector Machine (RBF) with Variable Selection — 2.59
  - 3. Support Vector Machine (Linear) — 2.70
- Oman (as of June 12, 2025)
  - 1. Support Vector Machine (Polynomial) with Variable Selection — 2.03
  - 2. Principle Component Regression with Variable Selection — 2.29
  - 3. Support Vector Machine (RBF) (oneSE) — 2.37
- Qatar (as of June 12, 2025)
  - 1. Support Vector Machine (RBF) (oneSE) — 2.36
  - 2. Random Forest (oneSE) — 3.44
  - 3. Multivariate Adaptive Regression Spline with Variable Selection — 3.77
- Saudi Arabia (as of April 23, 2025)
  - 1. Support Vector Machine (RBF) — 0.73
  - 2. Support Vector Machine (Polynomial) with Variable Selection — 1.14
  - 3. Multivariate Adaptive Regression Spline — 1.76
- United Arab Emirates (as of June 12, 2025)
  - 1. Random Forest with Variable Selection
  - 2. Random Forest (oneSE)

### Saudi Arabia Nowcasting and Forecast Accuracy (selected points)
- Nowcasts applied to real non-oil activities (excluding government activities) for comparability with official flash estimates.
- Data vintage for nowcasts: end-April 2025 (previous official series); official rebased GDP published in May 2025 not used for out-of-sample comparison to ensure fair comparison.
- Specific comparison for 2025Q1:
  - Official newly rebased GDP actual: 4.9 percent
  - Official flash estimate: 4.2 percent
  - Staff nowcast: 5.05 percent
  - Average forecast errors (percent, non-oil activities nowcasts):
    - IMF Staff Nowcasts (Avg forecast error -0.02pp)
    - Official Flash Estimates (Avg forecast error -0.14pp)

### Country-specific Indicator Highlights
- Bahrain: emphasis on POS transactions and CPI to capture domestic consumption and retail activity.
- Kuwait: unique indicators such as national valuation of gold; India’s intermediate goods production and PMI reflect trade linkages.
- Oman: reliance on external indicators due to limited domestic high-frequency data (imports, oil price inverse index, Saudi Arabia’s PMI, U.S. IP, U.S. Treasury yields).
- Qatar: CPI, imports, money supply, and external indicators (Dubai output, India PMI, U.S. consumer credit).
- Saudi Arabia: non-oil exports, CPI, cement deliveries, stock market indices, banking liquidity (3-month SIBOR, 4-week SAMA bill).
- UAE: international cargo/passenger traffic, industrial production indicators from U.S./U.K./Germany, and VIX.

### Robustness Checks, Sectoral Decomposition, and Validation
- Robustness tests: cross-validation, out-of-sample prediction, bottom-up decomposition, external composite index construction, external validation.
- Oman bottom-up decomposition:
  - Services GDP predictors: air transport, money supply (M2, quasi money), consumer prices, banking aggregates, U.S. interest/mortgage rates, regional monetary/credit aggregates.
  - Tradables GDP predictors: external trade data, banking indicators, monetary aggregates, regional indicators from Saudi Arabia.
  - Result: high alignment between aggregated subcomponent forecasts and overall non-oil nowcast; services account for significant volatility; tradables more sensitive to external shocks.
- External composite index validation:
  - Composite indices slightly reduce fine-grained explanatory power but maintain overall model performance.
  - Tradables GDP responds stronger to regional financial conditions; services GDP more affected by domestic liquidity and U.S. interest rates.

### Regional Comparison: Convergence and Divergence across GCC
- More diversified economies with stronger statistical systems (Saudi Arabia, UAE) show more robust nowcast performance due to richer indicators (financial markets, POS, trade variables).
- Narrower indicator sets (Oman, Kuwait) rely more on imports, credit aggregates, regional oil prices.
- Common predictors: oil prices, global PMI, regional equity indices, neighboring trade activity.
- GCC-specific challenges: sensitivity to oil production adjustments, fiscal swings, dependence on external capital flows, higher data latency, frequent national accounts revisions.
- Conclusion: tailored, country-specific nowcasting models with indicator flexibility and adaptive modeling are required.

### Operational Utility and Benchmarking
- Nowcasting model provides early signals of changes in non-oil momentum before incorporation into institutional or market forecasts.
- Direct statistical comparison with consensus forecasts limited by differences in frequency and target variable (quarterly non-oil GDP vs annual total GDP).
- Nowcasts generally aligned with directional trends in annual consensus projections for latter half of 2024.
- Framework serves as early-warning tool and complementary input into broader forecast discussions and institutional workflows.

### Conclusion and Policy Implications
- Framework strengths:
  - Integrates a wide range of high-frequency domestic and international indicators via structured filtering and transformation.
  - Model-agnostic methods balance flexibility and interpretability; Shapley decompositions enhance usability by identifying drivers of forecast revisions.
  - Iterative testing across >30 model variants and rigorous out-of-sample validation support credibility and generalizability.
  - Robustness checks confirm adaptability and policy relevance; in Saudi Arabia, nowcasts track and sometimes anticipate official flash estimates.
- Practical policy value:
  - Facilitates faster, data-driven policy decisions through timelier, sector-specific short-term signals.
  - Supports monitoring of growth momentum in non-oil sectors amid oil market volatility and diversification reforms.
  - Designed to be scalable and embedable into institutional workflows for oil-exporting and emerging market contexts.
- Future research avenues:
  - Integrate sentiment indices, mobility data, or supply chain disruption indicators to further enhance responsiveness.

### Annex I — Nowcasting Leading Indicators (selected retained indicators by country and sector)
- Bahrain (sample retained indicators):
  - Real: Bahrain: Consumer Price Index (NSA, Apr-19=100); Non-oil Imports (NSA, Thous.Dinars)
  - Oil: OPEC Reference Basket Price ($/BBL); Saudi Arabian Light: Spot Crude Price ($/BBL)
  - Monetary/Financial: Bahrain: Money Supply: M2 (SA, EOP, Mil.Dinars); Bahrain: Point of Sale Transaction Value (NSA, Dinars)
- Kuwait (sample retained indicators):
  - Real: Kuwait: Gold, National Valuation (EOP, Mil.US$)
  - Oil: Kuwait Export: Spot Crude Price ($/BBL)
  - Monetary/Financial: Kuwait: Credit Facility Agt. w Res: Consumer Loans (NSA, EOP, Mil.Dinars); Kuwait: Exchange Rate with US$ (EOP, Kuwaiti Dinar/US$)
- Oman (sample retained indicators):
  - Real: Imports: Electrical Machinery, Mechanical Equipment and Parts; Air Transport: Muscat International Airport: No of Flights: Abroad
  - Oil: U.S. Spot Oil Price: WTI; Oil Price Inverse Index Oman
  - Monetary/Financial: Market Capitalization MSX; U.S. 10 year Treasury Bond Yield
- Qatar (sample retained indicators):
  - Real: Qatar: Consumer Price Index (NSA, 2018=100); Dubai Total Economy Output (SA, 50+=Expansion)
  - Monetary/Financial: Money Supply: Currency in Circulation; Qatar Stock Exchange: shares Traded: Value: Real Estate
- Saudi Arabia (sample retained indicators):
  - Real: Saudi Arabia: Consumer Price Index (NSA, 2018=100); Saudi Arabia: Nonoil Exports (NSA, Mil.Riyals)
  - Oil: Crude Oil Production: Saudi Arabia (Thous. Barrels per Day)
  - Monetary/Financial: 4 Week SAMA Bill (Average, %); 3 Month SIBOR (Average, %)
- United Arab Emirates (sample retained indicators):
  - Real: Cargo Traffic: International; Passenger Traffic: International
  - Oil: European Brent Spot Price FOB ($/Barrel); United Arab Emirates Murban: Spot Crude Price ($/BBL)
  - Manufacturing: U.A.E. PMI: Total Economy (SA, 50+=Expansion); CBOE Market Volatility Index: VIX

### Annex II — Summary of Nowcasting Models (overview)
- Suite covers 22 models across:
  1. Linear and Regularized Regression Models: Step.Model, Elastic Net, Elastic Net (oneSE), with/without variable selection.
  2. Dimension Reduction Models: PCR, PCR with Variable Selection, PLS, PLS with Variable Selection.
  3. Tree-Based Models: MARS, Random Forest (and oneSE variants), Stochastic Gradient Boosting Trees (GBM and GBM oneSE), with/without variable selection.
  4. Support Vector Machines (SVMs): SVM (Linear/Polynomial/RBF) and oneSE or variable-selection variants.
- Models with variable selection trained on indicator subsets chosen via Shapley value decompositions or correlation screening.
- Assessment criterion: out-of-sample RMSE; alternative statistics used include MAE and MBE.

*IMF Working Paper — Nowcasting GCC GDP: A Machine Learning Solution for Enhanced Non-Oil GDP Prediction (Executive Summary and selected sections).*

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

### Executive Summary

### Overview
- Introduces a real-time nowcasting framework for estimating quarterly non-oil GDP growth in the GCC.
- Provides high-frequency, sector-specific estimates to give policymakers timely insight into GCC non-oil economic activities.
- Extends nowcasting literature for the MENA region by moving beyond single-model approaches and incorporating a richer set of high-frequency, cross-border indicators.
- First systematic application of a unified nowcasting framework across the GCC with country-specific models for non-oil GDP.

### Methodology and Framework
- Adopts a flexible, machine learning-based approach that incorporates a broad set of high-frequency indicators spanning real activity, financial conditions, trade, and oil-linked variables from domestic and global datasets.
- Uses a model-agnostic setup that encompasses 22 candidate models rather than a single-model specification (see Annex II for model list).
- Aggregates estimates across a suite of methods, including factor-based regressions, bridge equations, and machine learning techniques.
- Iterative design allows incorporation and reweighting of incoming information as new data are released and accommodates frequent data revisions and evolving statistical capacity in the GCC.

### Data and High-frequency Indicators
- Indicators include consumer prices and subcomponents, PMIs and industrial production, trade data, point-of-sale transactions, credit, deposit rates, and stock market indicators.
- Incorporates global, cross-border, and regional signals such as oil price benchmarks, interest rate spreads, and financial flows from key partners including the United States, China, Saudi Arabia, and the UAE.
- Tailors variable selection to capture spillover effects of oil market volatility on non-oil sectors and the impact of structural reforms aimed at diversification.

### Data Availability and Publication Lags (Selected Statistics)
- Quarterly GDP publication lags (from the reference quarter) and available time series (as of June 2025):
  - Bahrain: 2008Q1 – 2024Q4 +90 days
  - Kuwait: 2010Q1 – 2024Q4 +90 days
  - Oman: 2010Q1 – 2024Q4 +105 days
  - Qatar: 2011Q1 – 2024Q4 +90 days
  - Saudi Arabia: 2011Q1 – 2025Q1 +68 days
    - Note: Saudi Arabia authority (GASTAT) publishes flash estimates of quarterly GDP with a 30-day lag and the actual official data with a 68-day lag. The new rebased GDP published by GASTAT in May 2025 are not used for nowcasting in this paper, as we compare the out-of-sample performance against official estimates that were available at each respective point in time to ensure a fair comparison. Data vintage as of end-April 2025 was used for Saudi Arabia, reflecting chain-linked methodology adopted by GASTAT in early 2024.
  - United Arab Emirates: 2012Q1 – 2024Q3 +200 days [varies]

### Key Results and Contributions
- Produces high-frequency non-oil GDP nowcasts that enhance granularity, responsiveness, and transparency of short-term forecasts.
- Identifies key high-frequency indicators strongly correlated with non-oil activities in the GCC, helping to bridge long-standing information gaps.
- Complements existing institutional forecasts and, for Saudi Arabia, the model slightly outperformed official non-oil flash estimates in terms of sample average of forecast errors.

### Robustness, Evaluation, and Implementation
- Nowcasting framework includes model evaluation and selection procedures to ensure robustness across structural shifts and evolving data environments.
- Incorporates iterations with Shapley value decompositions to attribute contributions of indicators and models to nowcasts.
- Designed to be scalable and extendable across the GCC to strengthen economic surveillance and policy agility.

### Policy Implications and Practical Value
- Facilitates faster and data-driven policy decisions by improving the timeliness and sector-specificity of short-term economic signals.
- Helps authorities and analysts monitor the underlying growth momentum and recent trends in non-oil sectors amid oil market volatility and diversification reforms.
- Supports integration of improvements in national statistical systems in real time without the need to recalibrate models from scratch.

*Source: IMF Working Paper — Nowcasting GCC GDP: A Machine Learning Solution for Enhanced Non-Oil GDP Prediction (Executive Summary).*

### Section VI concludes.

### Section VI concludes.

### II. Literature Review — Parametric vs Non-Parametric Nowcasting Models
- Two broad classes of nowcasting models:
  - Parametric models:
    - Assume a specific functional form with a finite number of parameters.
    - Simplify estimation and inference but limit flexibility (Hastie, Tibshirani, & Friedman, 2009).
    - Examples cited: OLS (Bridge models), ARIMA/VAR (with mixed-frequency or Bayesian extensions), Dynamic Factor Models (DFM), State-Space Models (Giannone et al. 2008; Jensen et al. 2016; Nakazawa 2022; Almuzara et al. 2023).
  - Non-parametric models:
    - Do not assume a fixed functional form; offer flexibility useful when data structure is unknown or high-dimensional.
    - Require larger datasets, risk overfitting, and complicate interpretation and hypothesis testing (Breiman, 2001).
    - Examples cited: Random Forests, Support Vector Regression (SVR), Gradient Boosting (Ahmed et al. 2010; Fornaro et al. 2020; Richardson et al. 2021).

- MENA/GCC nowcasting evidence:
  - Alkhareif et al. (2022): generalized DFM for Saudi Arabia’s overall, oil, and non-oil real GDP growth.
  - Al-Rawashdeh (2024): Extreme Gradient Boosting for Jordan’s GDP; finds superior nowcast accuracy versus ARIMA and DFM.

- Study choice and rationale:
  - The paper adopts non-parametric machine learning algorithms to nowcast non-oil GDP using high-frequency indicators and several models.
  - Reasons: generally sufficient input data across GCC sectors; presence of dynamic interactions among indicators; potential to integrate more high-frequency/alternative data.
  - ML models used (Annex II): Support Vector Machines (SVMs), Random Forests, Stochastic Gradient Boosting Trees, Elastic Net, Principal Component Regression (PCR), and others.
  - Expected ML benefits: handle different data types and structures, mitigate multicollinearity and overfitting, improve prediction accuracy, and adapt continuously to new data.

### High-frequency Determinants (Table 2) — Sectors, Determinants, Expected Impact
- Sector groupings and listed determinants:
  - Real (incl. non-commodity trade):
    - Inflation; Imports and exports of non-commodity goods and services; Employment; PortWatch trade; Flights and hotels tracker.
    - Expected impact: Stable macroeconomic conditions, increased production/consumption/labor market activity foster growth and contribute to higher GDP.
  - Commodity Sector:
    - Volume of commodity trade; Price of commodity (energy, metal, agricultural products).
    - Expected impact: Higher prices and export volumes directly boost GDP; spillovers to non-commodity sector via fiscal and financial channels.
  - Manufacturing:
    - Industrial Production Index (IPI); Purchasing Managers’ Index (PMI).
    - Expected impact: Higher manufacturing outputs and business sentiment increase GDP.
  - Monetary and Finance:
    - Exchange rate; Interest rates; Point-of-sales; Credit growth; Stock market performance.
    - Expected impact: Influence trade competitiveness and inflation; loose financial conditions stimulate investment; higher stock indexes signal investor confidence.
  - Unconventional data sources:
    - Satellite imagery; Social media contents.
    - Expected impact: Signal production, consumption, and trade activity; increased activity indicates higher GDP.

- Notes and contextual points:
  - The distinction between domestic and global/regional variables is not specified; relevance depends on economic openness and data availability.
  - Point-of-sales data captures private consumption.
  - Unconventional sources mentioned: nighttime lights, land cover, nitrogen dioxide concentration, PortWatch trade, google search trends; relevance varies by country and structure.
  - Footnote: This study incorporates only a limited set of unconventional data sources because conventional indicators are not lacking for the GCC.

### III. Nowcasting Framework — Innovations and Components
- Two key innovations introduced:
  1. Advanced data collection and integration approach tailored to the GCC:
     - Broadens high-frequency indicators by combining traditional metrics with real-time financial transactions and unconventional data.
     - Uses advanced data scraping and partnerships with data providers to maintain a continuous, up-to-date pipeline.
     - Aims to improve granularity and responsiveness, narrowing the lag between developments and model outputs.
  2. Novel use of Shapley value decompositions:
     - Quantifies marginal contribution of each predictor (interpretability for Random Forests, Gradient Boosting Machines).
     - Used diagnostically and to iteratively refine indicator selection via a model-agnostic process.
     - Evaluates and reweights dozens of candidate indicators across more than 30 model variants to identify the most predictive and robust inputs for each country.
     - Produces a lean, country-calibrated, empirically validated set of high-frequency indicators.

- Main components of the framework (phased, modular approach):
  - Variable selection:
    - Rigorous process drawing from a wide range of high-frequency indicators covering domestic activity and external influences.
    - Establishes a dynamic information set for model accuracy.
  - Nowcasting stage:
    - Applies advanced ML algorithms; tests of 22 models (Annex II) are referenced to generate real-time estimates.
    - Algorithms are calibrated to maximize predictive accuracy and adapt to evolving data.
  - Shapley decompositions:
    - Assess marginal contribution of each predictor to nowcasts.
    - Enhance transparency, identify key drivers, and support iterative refinement of indicators and models.
    - Clarifying note: A Shapley value is the indicator’s contribution to the forecast; a negative Shapley value indicates the model interprets the recent movement as lowering predicted non-oil GDP regardless of the series’ direction.

- Intended outcome:
  - A flexible, data-rich framework combining methodological rigor, machine learning, and interpretability for real-time monitoring of non-oil GDP in GCC-like, resource-dependent economies.

### Variable Consideration and Selection — Data Pipeline and Selection Criteria
- Data extraction and filtering:
  - Millions of time series extracted from CEIC and Haver.
  - Filtered using Stata to retain approximately 1,000 country-relevant series per economy.
  - From the refined set, the top ~50 indicators are selected based on historical correlation with non-oil GDP growth.
  - Daily and weekly indicators converted to monthly frequency via aggregation or end-of-period selection.
  - Variables transformed into year-on-year growth rates for standardization across units and frequencies.

- Candidate evaluation core criteria:
  1. Predictive Capability:
     - Degree to which a transformed series exhibits historical correlation with non-oil GDP growth.
     - Typical threshold for inclusion: correlation coefficient greater than ±0.7.
     - Threshold adjusted based on data availability and correlation strength across countries.
  2. Historical Coverage:
     - Indicators expected to have at least 10 years of historical data to support robust model training and temporal variation for detecting predictive relationships.

- Additional methodological points:
  - The framework tests multiple configurations and country-specific dynamics to adapt indicator sets and preserve transparency and interpretability.

*IMF Working Paper — Nowcasting GCC GDP: A Machine Learning Solution for Enhanced Non-Oil GDP Prediction*

### 3. Timeliness (Latency) – Indicators must be released with minimal lag to ensure nowcasts reflect the

### 3. Timeliness (Latency) – Indicators must be released with minimal lag to ensure nowcasts reflect the most recent developments. On average, we target series with publication lags of no more than ±45 days.

### Timeliness and selection criteria
- Target publication lag: no more than ±45 days.
- Criteria applied iteratively:
  - Year-on-year transformation used for consistent evaluation across frequencies.
  - Correlation threshold is country-specific and based on how many candidate series exceed the statistical cut-off.
  - Thresholds chosen based on empirical testing to balance predictive power and data availability; thresholds are flexible with documented country-specific exceptions when justified by overall model performance.
  - Historical coverage exceptions are rare and only allowed when a variable adds material predictive power.
  - Indicators prioritized by latency if they provide sufficient lead time to predict the current quarter based on already-available information.
- Trade-off management:
  - All predictors must align temporally; the training sample is constrained by the start date of the series with the shortest history.
  - Adding predictors improves in-sample fit but may cause overfitting when sample size is limited.
  - Training window chosen to balance breadth of information with sufficient sample size.
  - Prior evidence: nowcasting models incorporating 10–15 indicators and spanning at least 10 years of data tend to perform best in accuracy and generalizability.
- Indicator basket maintenance:
  - Indicator basket typically locked for following iterations.
  - Revisit a series if Shapley/LIME diagnostics show two or more consecutive quarters of negligible contribution; test a sector-analogous substitute under the same cross-validation protocol.

### Data input description
- Nowcasting framework leverages high-frequency indicators capturing domestic and global dynamics:
  - Domestic: trade (imports and exports), financial market conditions (loans, credit growth, equity market performance), consumer prices, sector-specific indicators such as air transport.
  - Global: oil prices, international trade flows, macroeconomic policy uncertainty indices.
- Purpose:
  - Integrate high-frequency data across dimensions to provide comprehensive and timely view of current economic conditions.
  - Support real-time monitoring and evidence-based policymaking by uncovering data-driven patterns that inform short-term economic forecasts.

### Nowcasting model evaluation and selection
- Three-stage structured process:
  1. Training stage:
     - Train each model on historical dataset using a selected set of 10–15 high-frequency predictors and at least 10 years of data.
     - Target variable: non-oil real GDP.
     - Fit diverse set of algorithms to capture complex short-term relationships.
     - Multiple combinations of indicator sets tested iteratively over time to refine country-specific indicator sets.
  2. Holdout test stage:
     - Assess predictive performance on a holdout test set using Root Mean Squared Error (RMSE).
     - RMSE calculated as the square root of the average squared differences between predicted and observed values.
     - Emphasis on minimizing out-of-sample RMSE to avoid overfitting.
     - Preliminary filter: assess model and indicator combinations on a fixed holdout; surviving configurations move to stage three.
  3. Cross-validation and final selection:
     - Rolling window approach typically reserving the final 15 percent of the data as a holdout set and training on the initial 85 percent.
     - Over 30 model types are tested, each fine-tuned using the training set and evaluated on the holdout set.
     - Best-performing model based on out-of-sample RMSE is re-estimated using the full dataset and used to generate final nowcasts.
     - Models selected using alternative statistics (e.g., MAE and MBE) consistently produce closely aligned results.
- Algorithms employed (selected for their strengths):
  - Support Vector Machines (SVMs) for high-dimensional, linear and non-linear relationships.
  - Random Forests for robustness and handling large datasets.
  - Stochastic Gradient Boosting Trees to reduce risk of overfitting.
  - Elastic Net for high-dimensional and collinear features with automatic feature selection.
  - Principal Component Regression (PCR) to address multicollinearity by transforming predictors into uncorrelated variables.
- Model-agnostic approach: select most appropriate algorithm per country and data environment to ensure accuracy, adaptability, generalizability, and interpretability.

### Iterations with Shapley value decompositions
- Shapley values used to attribute predictive contribution of each indicator in multivariate models:
  - Quantify each variable’s marginal contribution to predictive accuracy.
  - Indicators with consistently low contributions are flagged for potential exclusion.
  - Cycle of model estimation, decomposition, and targeted refinement repeated across dozens of model structures and input configurations until convergence.
- Practical outcomes:
  - Each country’s final model often reflects more than 30 modeling variants and hundreds of predictor permutations.
  - Final tailored set: approximately 10–15 indicators per economy, each demonstrating consistent and substantial contribution as measured by Shapley values.
- Interpretation note:
  - Shapley values remain valid in presence of correlated variables; marginal contribution is shared among correlated variables and should be interpreted as shared predictive influence, not causal effects.
- Benefits:
  - Improves forecast precision, interpretability, and operational relevance for real-time surveillance.
  - Supports informed, timely, and targeted policy analysis.

### Nowcasting results for GCC — indicator selection and country-specific highlights
- Common indicators across the GCC:
  - Oil price indicators, including regional crude spot prices and the OPEC reference basket price.
  - Price and retail indicators: CPI and Point-of-Sales (POS) transactions.
  - Manufacturing and business sentiment indicators: domestic PMI and PMI of main trading partners or regional/global indices.
  - Monetary and financial sector variables: stock market indices, credit measures, and interest rate spreads.
- Country-specific indicator emphasis:
  - Bahrain:
    - Greater emphasis on POS transactions and the CPI due to importance of domestic consumption and retail activity.
  - Kuwait:
    - Distinct indicators such as national valuation of gold.
    - Inclusion of India’s intermediate goods production and India’s PMI highlights production and trade linkages.
  - Oman:
    - Reliance on external indicators due to limited high-frequency domestic data: import measures, oil price inverse index, Saudi Arabia’s PMI, U.S. industrial production, and U.S. Treasury yields.
  - Qatar:
    - CPI, imports, and money supply capture domestic consumption; external indicators such as Dubai’s total economy output, India’s PMI, and U.S. consumer credit capture regional and global demand spillovers.
  - Saudi Arabia:
    - Broad set beyond oil: non-oil exports, CPI, cement deliveries, stock market indices, banking system’s liquidity (proxied by 3-month SIBOR and 4-week SAMA bill).
  - UAE:
    - Emphasis on international cargo and passenger traffic, industrial production indicators from the U.S., U.K., and Germany, and the global volatility index (VIX).

### Nowcasting statistics — Top 3 performing models and Out-of-Sample RMSE (Q1 2025; forecasting non-oil GDP growth for Q2 2024)
- Bahrain (as of June 12, 2025)
  - 1. Stochastic Gradient Boosting Trees (oneSE) — 0.72
  - 2. Stochastic Gradient Boosting Trees — 1.07
  - 3. Random Forest (oneSE) with Variable Selection — 1.35
- Kuwait (as of June 12, 2025)
  - 1. Support Vector Machine (RBF) — 2.50
  - 2. Support Vector Machine (RBF) with Variable Selection — 2.59
  - 3. Support Vector Machine (Linear) — 2.70
- Oman (as of June 12, 2025)
  - 1. Support Vector Machine (Polynomial) with Variable Selection — 2.03
  - 2. Principle Component Regression with Variable Selection — 2.29
  - 3. Support Vector Machine (RBF) (oneSE) — 2.37
- Qatar (as of June 12, 2025)
  - 1. Support Vector Machine (RBF) (oneSE) — 2.36
  - 2. Random Forest (oneSE) — 3.44
  - 3. Multivariate Adaptive Regression Spline with Variable Selection — 3.77
- Saudi Arabia (as of April 23, 2025)
  - 1. Support Vector Machine (RBF) — 0.73
  - 2. Support Vector Machine (Polynomial) with Variable Selection — 1.14
  - 3. Multivariate Adaptive Regression Spline — 1.76
- United Arab Emirates (as of June 12, 2025)
  - 1. Random Forest with Variable Selection
  - 2. Random Forest (oneSE)

*IMF Working Papers — Nowcasting GCC GDP: A Machine Learning Solution for Enhanced Non-Oil GDP Prediction*

### 3.  Random Forest

### 3.  Random Forest

### Key model performance metrics
- 0.79
- 0.85
- 0.93
- Note: (1) We also include more traditional methods like OLS, which in some instances also perform well in comparison to the various machine learning models.  
- Note: (2) In addition to RMSE, model outputs provide alternative statistics of MAE and MBE. The top two to three models demonstrate robustness across various fitness statistics, consistently yielding closely aligned results.

### Nowcasting Estimations
- Out-of-sample predictions from training dataset: 2011Q1—2023Q1; testing dataset: 2023Q2-2024Q4.
- The panel presents historical nowcasting results for non-oil GDP growth across GCC countries, including actual outcomes and model-based estimates through Q1 2025.
- Nowcasts for Q1 of 2025 are as of the timline indicated in Table 3, April/June estimations, which have now been plotted against Q1 actuals available as of August 18, 2025.
- Saudi Arabia nowcasts:
  - Nowcasts are applied to real non-oil activities (excluding government activities) for consistency with official data releases and comparability with official flash estimates.
  - Saudi Arabia’s nowcasts results above are based on the previous official data published by GASTAT (data vintage as of end-April 2025) prior to the new rebased GDP in May 2025. The actual data shown above also refer to the previous official series, except for 2025Q1, which is only avilable under the new rebased GDP. See Figure 5 for comparison.
  - Note: (1) The official actual data refer to Saudi Arabia’s initial official release of quarterly real non-oil acticvities growh for each respective quarter. These figures may differ slightly from the current official data, which incorporate hisotrical revision (e.g., GASTAT’s methodolocal enhancement to national accounts statistic in 2024 and 2025).
  - Note: (2) The comparision for 2025Q1 should be interpreted with caution, as the official actual data (4.9 percent) is from GASTAT’s newly rebased GDP (published in May 2025), whereas the official flash estimates (4.2 percent) and staff nowcasts (5.05 percent) were based on the previous series.
  - Note: (3) Average forecast errors are calculated as the sample average of the projection minus the actual values, expressed in percentage points (pp).
- Forecast accuracy summary for Saudi Arabia (percent, non-oil activities nowcasts):
  - IMF Staff Nowcasts (Avg forecast error -0.02pp)
  - Official Flash Estimates (Avg forecast error -0.14pp)
  - Official Actual (at the time of publication)

### Regional Comparison: convergence and divergence across GCC
- Countries with broader diversification and stronger statistical systems (e.g., Saudi Arabia and the UAE) show more robust nowcast performance driven by:
  - financial market indicators
  - point-of-sale transactions
  - global trade variables
- Countries with narrower high-frequency indicator sets (e.g., Oman and Kuwait) rely more on:
  - import values
  - credit aggregates
  - regional oil prices
- Common regional predictors:
  - Oil prices remain important predictors across the GCC due to fiscal transmission channels and government (or quasi-government) led investment cycles.
  - Global PMI, regional equity indices, and trade activity in neighboring countries often serve as useful external signals.
- Cross-country differences:
  - Marginal predictive value of external variables varies significantly by country depending on structural reforms, diversification progress, and data maturity.
  - GCC economies generally benefit from more consistent macroeconomic reporting and a relatively wider set of high-frequency indicators compared to other MENA countries, especially on financial conditions.
  - GCC-specific challenges: more acute sensitivity to oil production adjustments, broader fiscal policy swings, heavier dependence on external capital flows, higher data latency, and more frequent historical revisions for national accounts.
- Conclusion: a one-size-fits-all approach is inadequate; tailored, country-specific nowcasting models with indicator flexibility and adaptive modeling are required.

### V. Robustness Checks and Forecasting Performance
- Robustness tests conducted include:
  - cross-validation
  - out-of-sample prediction
  - bottom-up decomposition
  - external composite index construction
  - external validation
- Purpose: ensure models are stable and adaptable across GCC economies given structural complexity and significant role of oil revenues.

### Bottom-Up Decomposition: Oman’s Non-Oil GDP
- Approach: nowcast services GDP and tradables GDP separately, then aggregate and compare to overall non-oil GDP nowcast.
- Services sector model predictors:
  - Air transport data (e.g., flight arrivals and departures)
  - Money supply metrics (M2 and quasi money)
  - Consumer prices
  - Financial aggregates such as government deposits and private sector liabilities
  - External variables: U.S. interest rates and mortgage rates
  - Monetary and credit aggregates from Saudi Arabia and the UAE
- Tradables sector model predictors:
  - External trade data (e.g., exports in INR and USD, imports of mineral products)
  - Financial system indicators (e.g., commercial bank deposits, resident bank balances)
  - Monetary aggregates related to private sector and foreign currency time deposits
  - Regional indicators from Saudi Arabia (PMI purchase prices, domestic claims, private credit) to account for regional spillovers
- Result:
  - High degree of alignment between aggregated subcomponent forecasts and the overall non-oil GDP nowcast, providing internal validation.
  - Services sector accounts for a significant portion of volatility and growth in Oman’s non-oil GDP.
  - Tradables sector shows greater sensitivity to external shocks and global demand conditions.

### Table 4. Oman Robustness Check Indicators (preserved structure and itemization)
- Tradable GDP Indicators
  - Exports:
    - Exports: INR: Oman
    - Exports: USD: Oman
  - Imports:
    - Imports: Mineral Products
  - Banking and Financial Indicators:
    - Deposits: Commercial Banks: Rial Omani: Over 2% to 3%
    - Commercial Banks: Assets: Balances Due from Banks: Resident
    - Deposits: Private Sector: Time: Foreign Currency
  - Saudi Arabia Financial Metrics:
    - Saudi Arabia PMI: Total Economy Purchase Prices
    - Saudi Arabia: Monetary Survey: Domestic Claims
    - Saudi Arabia: Private Sector Credit
  - Other Indicators:
    - Domestic Assets: Other Net Items
    - Number of Subscribers: Internet

- Services GDP Indicators
  - Air Transport: Muscat International Airport:
    - Num. of Flights: Abroad: Departures
    - Num. of Flights: Abroad: Arrivals
  - Banking and Money Supply:
    - Oman: Conventional Bank Liabilities: Government Deposits
    - Oman: Private Sector Deposits: Total
    - Oman: Money Supply: M2
    - Oman: Money Supply: Quasi Money
  - Inflation and Consumer Prices:
    - Oman: Consumer Prices
  - Interest Rates and Mortgage Rates:
    - U.S.: 5-Year Treasury Note Yield at Constant Maturity
    - U.S.: 30-Year Fixed Mortgage Rate
  - Saudi Arabia and UAE Financial Metrics:
    - Saudi Arabia: Monetary Base [Reserve Money]
    - UAE: Government Domestic Credit to Residents

### External Composite Index Validation
- Method: construct tailored external composite indices for Services GDP, Tradables GDP, and General Non-Oil GDP using international variables that consistently influenced model performance.
- Intended benefits:
  1. Model simplification: consolidate multiple external variables into a single composite input to make models more parsimonious and interpretable.
  2. Robustness validation: compare predictive accuracy of disaggregated external variables versus composite indices to assess sufficiency of aggregated signal.
  3. Sectoral sensitivity mapping: assess how each model responds to the composite index to identify how external conditions impact each segment.
- Findings:
  - Composite indices slightly reduce fine-grained explanatory power but maintain overall model performance for real-time tracking and streamlined policy monitoring.
  - Tradables GDP shows a stronger response to regional financial conditions than services GDP.
  - Services GDP is more affected by domestic liquidity and U.S. interest rate changes.

### Forecast Performance Benchmarking: Model vs. Consensus
- Direct statistical comparison with most consensus forecasts is not feasible because consensus forecasts are often annual and refer to total GDP, while the framework targets quarterly non-oil GDP.
- Qualitative benchmarking insights:
  - Model estimates for the latter half of 2024 are generally aligned with directional trends in annual consensus projections across all countries.
  - Nowcasting model can provide early signals of changes in economic momentum in the non-oil economy before incorporation into institutional or market forecasts.
  - In Saudi Arabia, nowcasts closely align with official flash estimates (GASTAT) and in several quarters have matched or slightly outperformed them in terms of final-revised accuracy.
- Operational utility:
  - Higher frequency and sectoral focus allow more timely and granular insights, complementing consensus forecasts that rely on slower-moving data and structural models.
  - Nowcasting framework serves as an early-warning tool and complementary input into broader forecast discussions.
  - Use of real-time nowcasts could help improve calibration and timeliness of consensus forecasts in data-constrained environments such as the GCC.

### VI. Conclusion
- The paper proposes a machine learning–based nowcasting framework tailored to GCC economies, focusing on quarterly non-oil GDP to deliver timely, granular, and sector-specific insights.
- Key strengths of the framework:
  - Integrates a wide range of high-frequency domestic and international indicators, transformed and filtered through a structured process.
  - Model-agnostic methods balance flexibility and interpretability and adapt to country-specific conditions while remaining transparent.
  - Shapley value decompositions enhance usability by identifying key drivers behind forecast revisions.
  - Iterative testing across more than 30 model variants and rigorous out-of-sample validation reinforce credibility, ensuring forecasts are accurate, reliable, and generalizable.
  - Robustness checks (sectoral decomposition for Oman and composite index test) confirm adaptability and policy relevance.
  - In Saudi Arabia, nowcasts track official flash estimates closely and sometimes anticipate revisions.
- Broader implications:
  - First region-wide application of machine learning-based nowcasting models tailored to GCC economic structures and data environments.
  - Introduces methodological innovations: automated, scalable data integration process and novel use of Shapley value decompositions.
  - Designed for operational utility and embedability into institutional workflows; scalable and replicable for monitoring real-time economic activity in oil-exporting and emerging market contexts.
- Future research avenues:
  - Integrate sentiment indices, mobility data, or supply chain disruptions to further enhance responsiveness of nowcasts.

*IMF Working Papers — Nowcasting GCC GDP: A Machine Learning Solution for Enhanced Non-Oil GDP Prediction*

### Annex I. Nowcasting Leading Indicators

### Annex I. Nowcasting Leading Indicators

### Purpose and framework
- Documents the leading indicators used in the nowcasting framework to predict quarterly non-oil GDP growth across the GCC.
- Indicators selected based on predictive relevance, frequency, and availability; span trade, prices, financial conditions, consumer behavior, and external variables.
- Inclusion reflects a balance between empirical correlation, economic intuition, and practical considerations such as data latency and update frequency.
- Underlying model diagnostics use Shapley value decomposition to gauge each variable’s marginal impact on the nowcast.
- A Shapley value is the indicator’s contribution to the forecast, not evidence that the series itself rose or fell (e.g., a negative Shapley value for broad money means the model interprets recent liquidity build-up as signaling inflationary or tightening pressures that lower the predicted non-oil GDP).

### Bahrain — Retained indicators (by sector)
- Real Sector
  - Bahrain: Consumer Price Index (NSA, Apr-19=100)
  - Bahrain: CPI: Recreation and Culture (NSA, Apr-19=100)
  - Bahrain: Non-oil Imports (NSA, Thous.Dinars)
- Oil Sector
  - OPEC Reference Basket Price ($/BBL)
  - Saudi Arabian Light: Spot Crude Price ($/BBL)
- Manufacturing
  - Global PMI: Composite Output (SA, 50+=Expansion)
  - Developed Markets PMI: Composite (SA, 50+=Expansion)
- Monetary and Financial
  - Bahrain: Public Debt: Conventional Instruments (EOP, NSA, Mil.Dinars)
  - Bahrain: Public Debt: Development Bonds (EOP, NSA, Mil.Dinars)
  - Bahrain: Money Supply: M2 (SA, EOP, Mil.Dinars)
  - Bahrain: Monetary Survey: Other Domestic Assets [Net] (NSA, EOP, Mil.Dinars)
  - Bahrain: One Month Deposit Rate (AVG, %)
  - Bahrain: Stock Exchange: All Share Index (EOP, Feb.02-Dec.02=1000)
  - Bahrain: Point of Sale Transaction Value (NSA, Dinars)
  - Bahrain: Point of Sale Transactions Number (NSA, Number)

### Kuwait — Retained indicators (by sector)
- Real Sector
  - Kuwait: Gold, National Valuation (EOP, Mil.US$)
- Oil Sector
  - Kuwait Export: Spot Crude Price ($/BBL)
- Manufacturing
  - India: IP: Intermediate Goods (NSA, Apr.11-Mar.12=100)
  - India PMI: Manufacturing Output Prices (NSA, 50+=Expansion)
- Monetary and Financial
  - Global excl Mainland China PMI: Services Future Activity (NSA, 50+=Expansion)
  - Saudi Arabia: General Share Price Index [TASI] (EOP, 1985=1000)
  - U.S.: Stock Price Index: NYSE Composite (EOP, Dec-31-02=5000)
  - U.S.: Commercial Bank Credit to the Private Sector (NSA, Bil.$)
  - China: Terms of Trade (NSA, 2010=100)
  - UAE: Deposits (EOP, NSA, Mil.AED)
  - Oman: Exchange Buying Rate: Kuwait (EOP, Omani Rial/Kuwaiti Dinar)
  - Kuwait: Credit Facility Agt. w Res: Consumer Loans (NSA, EOP, Mil.Dinars)
  - Kuwait: Boursa Kuwait Shares Traded Bought by Kuwaiti Co/Estab (EOP, Thous)
  - Kuwait: Exchange Rate with US$ (EOP, Kuwaiti Dinar/US$)
  - Kuwait: Time Deposits of Res & Non-Res: Over 1-3 months (NSA, EOP, Mil)

### Oman — Retained indicators (by sector)
- Real Sector
  - Imports: Electrical Machinery, Mechanical Equipment and Parts
  - Imports: Food and Food Preparation Beverages and Tabacco
  - Air Transport: Muscat International Airport: No of Flights: Abroad
  - Bank Clearance: No of Items
  - Consumer Price Index
  - Exports: SITC: Middle East and North Africa: Oman
- Oil Sector
  - U.S. Spot Oil Price: WTI
  - Oil Price Inverse Index Oman
- Manufacturing
  - Saudi Arabia PMI: Total Economy Output
  - U.S. Industrial Production excluding Construction
- Monetary and Financial
  - Loans: Commercial Banks: Over 0-5%
  - Market Capitalization MSX
  - MSX Index: Banking and Investment
  - U.S. 10 year Treasury Bond Yield
  - U.S. News Based Economic Policy Uncertainty Index

### Qatar — Retained indicators (by sector)
- Real Sector
  - Qatar: Consumer Price Index (NSA, 2018=100)
  - Dubai Total Economy Output (SA, 50+=Expansion)
  - International Rig Count: Land: Qatar
  - Import Price: Crude Oil, Qatar
- Manufacturing
  - India PMI: Services Employment
- Monetary and Financial
  - Imports: Qatar
  - Deposit Rate: Weighted Average: 3 months
  - Money Supply: Currency in Circulation
  - Qatar Stock Exchange: shares Traded: Value: Real Estate
  - Liabilities: Banks: Domestic: Due to Qatar Central Bank
  - US Consumer Credit Outstanding
  - Saudi Arabia Imports

### Saudi Arabia — Retained indicators (by sector)
- Real Sector
  - Saudi Arabia: Consumer Price Index (NSA, 2018=100)
  - Saudi Arabia: Nonoil Exports (NSA, Mil.Riyals)
- Oil Sector
  - Crude Oil Production: Saudi Arabia (Thous. Barrels per Day)
  - Saudi Arabian Light: Spot Crude Price ($/BBL)
- Manufacturing
  - Global Manufacturing PMI (SA, 50+=Expansion)
  - Saudi Arabia: Cement Deliveries (NSA, Thous.Ton)
- Monetary and Financial
  - 4 Week SAMA Bill (Average, %)
  - 3 Month SIBOR (Average, %)
  - Stock Market: Number of Transactions (Number)
  - Stock Market: Number of Shares Traded (Mil)
  - Stock Market: Tadawul All Share Index (EOP, 1985=1000)
  - Foreign Reserves: Investment in Foreign Securities (EOP, Mil.Riyals)

### United Arab Emirates — Retained indicators (by sector)
- Real Sector
  - Cargo Traffic: International
  - Passenger Traffic: International
- Oil Sector
  - European Brent Spot Price FOB ($/Barrel)
  - Dubai Fateh: Spot Crude Price ($/BBL)
  - United Arab Emirates Murban: Spot Crude Price ($/BBL)
- Manufacturing
  - U.A.E. PMI: Total Economy (SA, 50+=Expansion)
  - Global PMI: Composite Output (SA, 50+=Expansion)
  - Global PMI: Composite New Orders (SA, 50+=Expansion)
  - Global PMI: Services Business Activity (SA, 50+=Expansion)
  - Saudi Arabia PMI: Total Economy New Orders (NSA, 50+=Expansion)
  - Saudi Arabia PMI: Total Economy (NSA, 50+=Expansion)
  - UK: Index of Production (SA, 2019=100)
  - U.S.: Industrial Production excluding Construction (NSA, 2017=100)
  - Germany: Industrial Production: Total Industry incl Construction (NSA, 2015=100)
  - CBOE Market Volatility Index: VIX
  - CBOE Vix Volatility Index

_Annex I. Nowcasting Leading Indicators — IMF Working Paper content excerpt_

### 4. Monetary and Financial

### 4. Monetary and Financial

### Market volatility indicators
- CBOE Market Volatility Index: VIX (Index)
  - A real-time market index representing the market's expectation of 30-day forward-looking volatility.
- CBOE VIX Volatility Index [VVIX] (AVG, Index)
  - Represents the volatility of the VIX itself, providing a measure of the market's expectation of the future volatility of the VIX.
- Note: The indicators selected are part of an exhaustive list that spans all sectors and available data. For example, in the case of Qatar, natural gas indicators were tested but not chosen given their performance in the models.

### Annex II — Summary of Nowcasting Models (overview)
- This annex provides a comprehensive overview of the 22 models evaluated in the nowcasting framework.
- The selection spans linear, regularized, dimension-reduction, tree-based, and kernel-based machine learning algorithms, applied both with and without variable selection techniques.
- Models marked “(oneSE)” follow the one-standard-error rule for optimal complexity.
- All models were assessed based on out-of-sample Root Mean Squared Error (RMSE) to determine predictive performance.
- Each model was tested on a consistent training sample with harmonized input variables.
- Models with variable selection were trained on indicator subsets chosen via Shapley value decompositions or correlation screening.
- Full hyperparameter tuning procedures and model specifications are available upon request.

### 1. Linear and Regularized Regression Models
- Step.Model: Stepwise regression using forward, backward, or both directions to select features.
- Elastic Net: Regularized linear model combining L1 and L2 penalties to manage collinearity.
- Elastic Net (oneSE): Simpler Elastic Net variant selected via cross-validation using the one-standard-error rule.
- Elastic Net with Variable Selection / Elastic Net (oneSE) with Variable Selection: Elastic Net models incorporating pre-screening of variables based on predictive importance.

### 2. Dimension Reduction Models
- Principal Component Regression (PCR): Regression using principal components of the predictors.
- PCR with Variable Selection: PCR applied to a subset of variables selected prior to transformation.
- Partial Least Squares Regression (PLS): Dimension-reduction technique that maximizes covariance between predictors and target.
- PLS with Variable Selection: PLS model applied to a reduced set of predictors.

### 3. Tree-Based Models
- Multivariate Adaptive Regression Splines (MARS): Nonlinear regression that models interactions and nonlinearities via piecewise linear basis functions.
- MARS with Variable Selection: MARS model applied to a filtered subset of predictors.
- Random Forest: Ensemble of decision trees using bootstrap aggregation and random feature selection.
- Random Forest (oneSE): Random Forest with a reduced number of trees selected under the one-standard-error rule.
- Random Forest with Variable Selection / Random Forest (oneSE) with Variable Selection: Random Forests trained on a selected set of predictors.
- Stochastic Gradient Boosting Trees (GBM): Boosting method that sequentially fits shallow trees to residuals.
- GBM (oneSE): Parsimonious boosting model tuned using the one-standard-error rule.

### 4. Support Vector Machines (SVMs)
- SVM (Linear): Linear-kernel SVM used for regression (SVR).
- SVM (Linear) (oneSE): Linear SVM with regularization selected using the one-standard-error rule.
- SVM (Polynomial): SVR using polynomial kernel to model nonlinear relationships.
- SVM (Polynomial) (oneSE): Polynomial-kernel SVR with reduced complexity.
- SVM (RBF): SVR with radial basis function kernel for capturing nonlinearities.
- SVM (RBF) (oneSE): Simpler variant of RBF-kernel SVR.
- SVM (Linear/Polynomial/RBF) with Variable Selection: Kernel-based models trained on pre-selected variable sets.

### Modeling and implementation notes
- Models with variable selection were trained on indicator subsets chosen via Shapley value decompositions or correlation screening.
- Assessment criterion: out-of-sample Root Mean Squared Error (RMSE).
- The suite of models covers a broad set of algorithmic approaches to address limited data span and nonlinearities in nowcasting GCC non-oil GDP.

*Nowcasting GCC GDP: A Machine Learning Solution for Enhanced Non-Oil GDP Prediction — Working Paper No. WP/2025/268*

---


_Source: https://www.imf.org/-/media/files/publications/wp/2025/english/wpiea2025268-source-pdf.pdf_