## wpiea2026108-source-pdf

## Source details

**Canonical URL:** [wpiea2026108-source-pdf](https://www.imf.org/-/media/files/publications/wp/2026/english/wpiea2026108-source-pdf.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/wp/2026/english/wpiea2026108-source-pdf.pdf.md)
- [Structured JSON version](/-/media/files/publications/wp/2026/english/wpiea2026108-source-pdf.pdf.json)

---

### Executive Summary — Overview
- Access to reliable and timely data is critical for policy making; non-traditional data can generate accurate, real-time estimates ahead of official releases.
- Non-traditional data types used: indices extracted from satellite imagery, Google Trends, and flight tracking information (FlightRadar24).
- Applied to: (i) agricultural gross value added (GVA), (ii) tourism revenues, and (iii) the unemployment rate.

### Timeliness and Publication/Nowcast Lags
- Agricultural GVA:
  - Publication lag: ~t+90 days (1 quarter after the end of the reference quarter)
  - Nowcast lag: ~t-50 days (40 days into the reference quarter)
  - Practical timing: estimates generated starting around one month into the quarter, up to around five months before the official release (official release available 3 months after the end of the quarter).
- Tourism Revenues:
  - Publication lag: ~t+30 days (1 month after the end of the reference month)
  - Nowcast lag: ~t+5 days (5 days after the end of the reference month)
  - Practical timing: estimates generated 5 days after the month, around 3 weeks in advance of official figures (official figures available 4 weeks after the end of the month).
- Unemployment Rate:
  - Publication lag: ~t+30 days (1 month after the end of the reference quarter)
  - Nowcast lag: ~t-60 days (30 days into the reference quarter)
  - Practical timing: estimates generated starting one month into the quarter, up to three months ahead of the official release (official release available one month after the end of the quarter).

### Key Findings and Performance Metrics
- Inclusion of non-traditional, high-frequency indicators into nowcasting models yields substantial improvements in forecast accuracy across all three indicators.
- Out-of-sample RMSE reductions relative to benchmark ARIMA:
  - Agricultural GVA: 18 percent
  - Tourism revenues: 14 percent
  - Unemployment rate: 25 percent
- Informational value of non-traditional data evidenced by consistent outperformance of enriched models versus benchmarks.
- Whether machine learning or statistical models achieve higher predictive accuracy depends on the series and sector dynamics.

### Sector-specific Applications and Relevance
- Agriculture GVA:
  - Satellite-based indicators used: vegetation and climate indices.
  - Morocco exposure: drought risk; food prices account for 39 percent of the inflation basket; drought affects rural employment and economic stability.
- Tourism Revenues:
  - Predictors: travel-related Google Trends query data and flight tracking (FlightRadar24).
  - Context: tourism is a key growth driver; expanded air connectivity and infrastructure upgrades; major events include AFCON 2025 and the 2030 FIFA World Cup.
- Unemployment Rate:
  - Predictors: Google Trends job-related search queries.
  - Policy target: Job Plan 2030 aims to reduce unemployment rate to 9 percent by 2030.
  - Dynamics: job losses in agriculture driven by drought have weighed on employment; youth unemployment rising.

### Methodological and Practical Notes (end-to-end pipeline)
- Matching non-traditional indicators to target dynamics:
  - Agriculture: vegetation and climate indices from satellite imagery.
  - Tourism: Google Trends + FlightRadar24 passenger capacity.
  - Unemployment: Google Trends job-search indices.
- Pipeline components: data extraction, pre-processing, variable selection, model architecture; designed as a replicable framework.
- Models compared: machine learning families and statistical time-series families; performance varies by sector.
- Generalizability: scalable given near-global coverage of data sources.

### Policy Implications and Scalability
- Near-real-time, granular non-traditional data can reduce informational lags, especially during external shocks or turning points.
- Low-income and data-scarce economies: non-traditional data can serve as proxies where official data are unavailable.
- Emerging markets and advanced economies: provide early estimates and time savings for policymakers.
- Transferability: methods and data sources readily transferable due to near-global coverage and minimal replication requirements.

### Literature Review — Cross-cutting findings and sectoral applications
- Non-traditional data increasingly used for nowcasting; examples include satellite indicators, AIS vessel tracking, large-scale textual data integration, Google Places/Trends, and weekly trackers.
- Sectoral literature:
  - Agriculture: country studies cited (e.g., Colombia, India).
  - Tourism: many studies focus on arrivals and overnight stays; fewer on revenue-based measures.
  - Labor market: Google Trends well established for labor market nowcasting in advanced economies, fewer applications in emerging markets.
- Satellite/remote-sensing technical properties and timeliness (as reported):
  - Sentinel-2: resolutions as fine as 10 meters.
  - MODIS vegetation index: composited into 16-day products.
  - Sentinel-2 revisit period: approximately five days.
- Widely used satellite indicators and roles:
  - Agriculture Stress Index (ASI): dekadal; monthly summary; used as early warning for drought and recovery.
  - Precipitation: proxy for water availability; contrasts highlighted between January 2025 (drought) and January 2026 (rainfall recovery).
  - NDVI, EVI, VHI, NDVI Anomalies, Temperature (air at 2 m), Nighttime lights, Groundwater Storage Percentile.

- Summary table of satellite indicators (frequency, timeliness, source) — timeliness figures are approximate:
  - ASI: Frequency = Dekadal (10 days); Timeliness = near real-time; Source = Agricultural stress index near real time (Global - Dekadal - 1 km ) - FAO ASIS
  - Precipitation: Frequency = Daily; Timeliness = ~5-6 days; Source = ERA5-Land Daily Aggregated - ECMWF Climate Reanalysis
  - NDVI: Frequency = 8 days; Timeliness = ~1-2 weeks; Source = Landsat Collection 2 Tier 1 Level 2 8-Day NDVI Composite
  - EVI: Frequency = 8 days; Timeliness = ~1-2 weeks; Source = Landsat Collection 2 Tier 1 Level 2 8-Day EVI Composite
  - VHI: Frequency = Dekadal (10 days); Timeliness = near real-time; Source = Vegetation health index near real time (Global - Dekadal - 1 km) - FAO ASIS
  - NDVI Anomalies: Frequency = Dekadal (10 days); Timeliness = near real-time; Source = Non-Seasonal Indicators (Crops - Agricultural Stress Index System)
  - Temperature: Frequency = Daily; Timeliness = ~5-6 days; Source = ERA5-Land Daily Aggregated - ECMWF Climate Reanalysis
  - Nighttime Lights: Frequency = Daily; Timeliness = ~1-2 weeks; Source = VNP46A2: VIIRS Lunar Gap-Filled BRDF Nighttime Lights
  - Groundwater Storage Percentile: Frequency = Weekly; Timeliness = ~2-9 days; Source = Groundwater storage percentile (GRACE-DA-DM Global v3.0, 0.25°, weekly)

### High-frequency proxies: FlightRadar24 and Google Trends
- FlightRadar24 / Passenger Capacity (IMF Statistics Department indicator):
  - Daily indicator; available with a 5-day lag.
  - Estimates total seating capacity by aircraft type.
  - Correlation with official tourism arrivals: 0.9 in levels and 0.95 when smoothed with a 3-month moving average.
  - Disaggregated by country and airport; captures events (increase between December 2025 and January 2026 coinciding with AFCON 2025).
  - Official arrivals and tourism revenues: released with a 1-month lag.
- Google Trends:
  - Relative search intensity scaled 0–100; available at daily/weekly/monthly frequencies.
  - Timeliness: near real-time at daily frequency.
  - Normalization: index represents relative search interest, not absolute volumes.
  - Tourism revenues and unemployment show notable associations with multiple Google Trends indices in different languages.

### Methodology — Data transformation and processing
- Targets: growth rates or levels from official statistics; predictors: high-frequency indicators aggregated to target frequency.
- Aggregation: sum for flow variables, mean for indices; expressed in YoY growth rates where applicable.
- Temporal vintages: contemporaneous quarter t and lags t-1 and t-2 (where specified).
- Pre-estimation filtering: remove highly correlated variables at 90 percent threshold; filter indices with excessive zeroes.
- Agriculture Value Added specifics:
  - Target: YoY growth rate of quarterly agricultural GVA in volumes.
  - Predictors: daily/weekly/dekadal satellite-derived indicators aggregated to quarterly frequency and expressed in YoY growth rates.
  - Temporal vintages: t, t-1, t-2.
- Tourism Revenues specifics:
  - Target: monthly tourism receipts.
  - Predictors: monthly Google Trends indices and daily Passenger Capacity aggregated to monthly frequency.
  - Google Trends extraction date: April 3, 2026 (fixed sample window).
  - Transformations: Passenger Capacity = log-difference; Google Trends = log(1+p).
  - Query set: 100 tourism-related search terms in French, Spanish, and English.
  - Lag evaluation: Google Trends indices evaluated from t to t-6.
- Unemployment Rate specifics:
  - Target: quarterly unemployment rate.
  - Predictors: monthly Google Trends indices (100 search terms in French, Arabic, and English) aggregated to quarterly frequency by mean.
  - Google Trends extraction date: April 3, 2026.
  - Lag evaluation: contemporaneous (t) to two quarters (t-2).
  - Structural break: dummy for change since Q4 2019 (COVID).

### Nowcasting framework — Models, tuning, evaluation
- Twelve model specifications across three families: linear, tree-based ensemble, and time series.
- Variable selection paired with models to address high dimensionality.
- Benchmark: univariate ARIMA on the target alone.
- Models explored include: Random Forest (with and without variable selection), XGBoost (with and without variable selection), Elastic Net (with and without variable selection), PLS (with and without variable selection), OLS (with and without variable selection), ARIMAX (with and without variable selection), and ARIMA.
- Variable selection methods: MDI for RF, gain-based for XGBoost, Elastic Net non-zero coefficients, VIP for PLS, LASSO selection for OLS_VS, SIS + PCA for ARIMAX_VS.
- Hyperparameter tuning: training set 80 percent; five-fold time-series cross-validation; one-standard-error rule for complexity.
- Performance metrics: Primary = RMSE; Complementary = MAE; relative performance against ARIMA benchmark.
- Evaluation: rolling pseudo-out-of-sample with expanding estimation window.
- Train/test coverage and nowcast targets:
  - Agricultural GVA: Training = 2015Q1 - 2023Q3; Testing = 2023Q4 - 2025Q4; Nowcast = 2026Q1
  - Tourism Revenues: Training = 2019M11 - 2024M10; Testing = 2024M11 - 2026M2; Nowcast = 2026M3
  - Unemployment Rate: Training = 2006Q1 - 2021Q4; Testing = 2022Q1 - 2025Q4; Nowcast = 2026Q1
- Robustness checks:
  - Sub-sample stability (split test period in two).
  - Iterative updates (framework updated as new observations arrive).
  - Statistical significance: Clark–West test on pseudo out-of-sample forecasts.

### Results — Overview
- All twelve models estimated and compared via rolling pseudo-out-of-sample RMSE.
- Agriculture GVA: Random Forest delivers the lowest out-of-sample error.
- Tourism Revenues: best model is ARIMAX with PCA-based dimensionality reduction + passenger capacity as exogenous regressor.
- Unemployment Rate: best model is ARIMAX with PCA-based dimensionality reduction + structural break dummy.
- ARIMAX_VS outperforms ARIMAX with full Google Trends set for tourism and unemployment.
- Clark–West test indicates statistically significant improvements at the 5 percent level over the benchmark across all three indicators (interpret with caution given short evaluation windows).
- Broad implication: choose model family based on data nature — machine learning for complex nonlinear satellite indicators; statistical time-series for structured dynamics with exogenous inputs.

### Results — Agriculture Value Added (selected results)
- Final predictor set: EVI, NDVI anomalies, temperature, precipitation, groundwater storage percentile.
- Random Forest performance:
  - RMSE improvement over ARIMA: 18 percent
  - RF — Tree-based — RMSE = 2.82 — MAE = 2.35
  - ARIMA — Time series — RMSE = 3.46 — MAE = 2.68
  - Other reported model values (model — family — RMSE — MAE):
    - ARIMAX_VS — Time series — 5.06 — 3.41
    - RF_VS — Tree-based — 6.07 — 4.33
    - ARIMAX — Time series — 6.20 — 5.46
    - XGB — Tree-based — 6.94 — 6.44
    - XGB_VS — Tree-based — 6.94 — 6.44
    - EN — Linear — 8.03 — 5.18
    - PLS_VS — Linear — 8.08 — 6.51
    - PLS — Linear — 9.20 — 7.68
    - EN_VS — Linear — 9.26 — 6.93
    - OLS — Linear — 12.39 — 10.77
    - OLS_VS — Linear — 12.39 — 10.77
- RF captures turning points including rebound in early 2026; Shapley decomposition: vegetation signals dominant, followed by temperature and groundwater storage percentile.

### Results — Tourism Revenues (selected results)
- Contemporaneous Google Trends signals outperform lagged specifications.
- Best model: ARIMAX_VS with PCA retaining two principal components and passenger capacity as exogenous regressor.
- Performance:
  - RMSE improvement relative to ARIMA: 14 percent
  - ARIMAX_VS — Time series — RMSE = 0.19 — MAE = 0.14
  - ARIMA — Time series — RMSE = 0.22 — MAE = 0.16
  - Other reported model values:
    - RF — Tree-based — 0.22 — 0.17
    - RF_VS — Tree-based — 0.23 — 0.18
    - XGB — Tree-based — 0.28 — 0.21
    - XGB_VS — Tree-based — 0.28 — 0.21
    - PLS — Linear — 0.32 — 0.25
    - PLS_VS — Linear — 0.43 — 0.35
    - EN_VS — Linear — 0.46 — 0.38
    - EN — Linear — 0.48 — 0.39
    - OLS_VS — Linear — 0.55 — 0.44
    - OLS — Linear — 2.53 — 1.44
    - ARIMAX — Time series — 2.73 — 2.04
- ARIMAX estimated coefficients: passenger capacity coefficient = 0.5 (largest coefficient); Google Trends principal components add incremental predictive power.

### Results — Unemployment Rate (selected results)
- Contemporaneous Google Trends signals outperform lagged specifications.
- Best model: ARIMAX_VS with PCA retaining six principal components + COVID structural break dummy.
- Performance:
  - RMSE improvement relative to ARIMA: 25 percent
  - ARIMAX_VS — Time series — RMSE = 0.44 — MAE = 0.31
  - ARIMA — Time series — RMSE = 0.59 — MAE = 0.45
  - Other reported model values:
    - RF_VS — Tree-based — 0.69 — 0.55
    - RF — Tree-based — 0.72 — 0.58
    - OLS_VS — Linear — 0.79 — 0.68
    - PLS_VS — Linear — 0.85 — 0.74
    - XGB — Tree-based — 0.92 — 0.74
    - XGB_VS — Tree-based — 0.92 — 0.74
    - PLS — Linear — 1.05 — 0.95
    - EN — Linear — 1.64 — 1.50
    - EN_VS — Linear — 1.73 — 1.60
    - OLS — Linear — 1.99 — 1.51
    - ARIMAX — Time series — 2.20 — 1.86
- PCA aggregates search behavior across fr, ar, en into components related to applications, resignations, internships, unemployment; loadings indicate direction/magnitude but not causality.

### Conclusions and Next Steps
- Non-traditional data can meaningfully complement official statistics to provide timelier and more accurate nowcasts for Morocco (agriculture, tourism, unemployment).
- Models incorporating non-traditional data consistently surpass ARIMA benchmarks.
- No single method dominates across all indicators; model choice depends on sector and data type.
- Framework is designed for replicability across countries and sectors (industry, construction, retail).
- Future work: explore additional non-traditional data types, alternative ML and statistical configurations, and deployment across data-scarce countries to aid policymakers.

*Source: IMF Working Paper No. WP/2026/108, Section 5 and concluding remarks.*

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

### Executive Summary

### Overview
- Access to reliable and timely data is critical for the quality of policy making. Non-traditional data sources can generate accurate, real-time estimates of economic conditions, providing policymakers and analysts with timely data-driven insights ahead of official statistical releases.
- Non-traditional data types used: indices extracted from satellite imagery, Google Trends, and flight tracking information (FlightRadar24).
- The approach is applied to: (i) agricultural gross value added (GVA), (ii) tourism revenues, and (iii) the unemployment rate.

### Timeliness and Publication/Nowcast Lags
- Agricultural GVA:
  - Publication lag: ~t+90 days (1 quarter after the end of the reference quarter)
  - Nowcast lag: ~t-50 days (40 days into the reference quarter)
  - Practical timing: estimates generated starting around one month into the quarter, which is up to around five months before the official release (official release available 3 months after the end of the quarter).
- Tourism Revenues:
  - Publication lag: ~t+30 days (1 month after the end of the reference month)
  - Nowcast lag: ~t+5 days (5 days after the end of the reference month)
  - Practical timing: estimates generated 5 days after the month, which is around 3 weeks in advance of official figures (official figures available 4 weeks after the end of the month).
- Unemployment Rate:
  - Publication lag: ~t+30 days (1 month after the end of the reference quarter)
  - Nowcast lag: ~t-60 days (30 days into the reference quarter)
  - Practical timing: estimates generated starting one month into the quarter, which is up to three months ahead of the official release (official release available one month after the end of the quarter).

### Key Findings and Performance Metrics
- Inclusion of non-traditional, high-frequency indicators into nowcasting models yields substantial improvements in forecast accuracy across all three indicators studied.
- Out-of-sample Root Mean Squared Error (RMSE) reductions relative to benchmark models that rely solely on conventional time-series methods:
  - Agricultural GVA: 18 percent
  - Tourism revenues: 14 percent
  - Unemployment rate: 25 percent
- The informational value of non-traditional data is evidenced by consistent outperformance of enriched models versus benchmarks across the three indicators.
- Whether machine learning models or statistical models achieve higher predictive accuracy depends on the nature of the underlying series and the specific dynamics of each sector.

### Sector-specific Applications and Relevance
- Agriculture GVA:
  - Satellite-based indicators (vegetation and climate indices) are combined to capture drivers of agricultural activity.
  - Morocco’s agricultural sector is exposed to drought risk, with direct consequences for food prices—food prices account for 39 percent of the inflation basket—rural employment, and overall economic stability.
- Tourism Revenues:
  - Travel-related Google Trends query data and flight tracking information from FlightRadar24 are used to monitor tourism activity.
  - Tourism is a key growth driver; Morocco has expanded air connectivity, upgraded infrastructure, and prepared to host major events including the 2025 Africa Cup of Nations (AFCON) and the 2030 FIFA World Cup.
- Unemployment Rate:
  - Google Trends search query data capture labor market dynamics.
  - Job Plan 2030 targets a reduction in the unemployment rate to 9 percent by 2030.
  - Job losses in agriculture driven by drought have weighed on employment, particularly in rural areas, while youth unemployment has continued to rise.

### Methodological and Practical Notes
- The framework matches non-traditional indicators to the dynamics of the underlying series:
  - Agriculture: vegetation and climate indices derived from satellite imagery.
  - Tourism: Google Trends search queries and flight tracking data.
  - Unemployment: Google Trends search queries related to labor market dynamics.
- The end-to-end pipeline includes data extraction, pre-processing, variable selection procedures, and model architecture, designed as a replicable framework that can be tailored to different countries and sectors.
- The paper compares machine learning models and statistical models, finding performance varies by sector and data characteristics.
- The nowcasting framework is generalizable and scalable given the near-global coverage of the non-traditional data sources used.

### Policy Implications and Scalability
- Near-real-time, granular non-traditional data can reduce informational lags that constrain policy analysis, particularly during external shocks or turning points in the business cycle.
- For low-income and data-scarce economies, non-traditional data can serve as proxies for economic activity where official data is unavailable.
- For emerging markets and advanced economies, these sources provide early estimates and time savings in generating estimates for policymakers.
- The methods and data sources are readily transferable to other countries due to near-global coverage and minimal additional requirements for replication and adaptation.

*Source: Executive Summary, IMF Working Paper "Leveraging Non-traditional Data for Macroeconomic Nowcasting: The Case of Morocco".*

### 2. Literature Review

### 2. Literature Review

### Use of Non-traditional Data for Nowcasting: Cross-cutting Findings
- Non-traditional data are increasingly used in monitoring economic activity and nowcasting models to produce timely estimates of macroeconomic indicators.
- Examples of applications in the literature:
  - Fotopoulou et al. (2026): Combine satellite-derived indicators (nighttime lights, NDVI, EVI, nitrogen dioxide emissions) with traditional macro data in a Random Forest nowcasting framework to address near-complete absence of official GDP data in Venezuela.
  - Arslanalp et al. (2025): Nowcasting model for global maritime trade using satellite-based Automatic Identification System (AIS) data to track vessel movements in near real-time.
  - Zheng et al. (2024): Nowcast GDP growth rate and inflation expectation in China by integrating traditional macro data with large-scale textual data.
  - Austin et al. (2021): Combine Google Places API and Google Trends to develop high-frequency indicators aligned with official statistical concepts and classifications.
  - Woloszko (2020): OECD Weekly Tracker producing real-time weekly GDP growth estimates across 46 countries using short-history alternative data like Google Trends.
- Sectoral applications highlighted:
  - Agriculture: Bravo Higuera et al. (2024) (Colombia), Kaustubh et al. (2024) (India).
  - Tourism: Crispino and Mariani (2025) (payment card data + Google Search for Italy), Bespalova (2022) (Aruba; ARIMAX + Google Trends), Havránek and Zeynalov (2021) (Prague), Cevik (2020) (The Bahamas).
  - Labor market: Costa et al. (2024) (Portugal; daily Google Trends + MIDAS), Borup and Schütte (2022) (U.S.; Random Forest with Google Trends), Caperna et al. (2022) (EU27; Google Topics + Random Forest with Boruta), D’Amuri and Marcucci (2017) (U.S.; Google index GI).

### Gaps and Scope for Extension
- Many applications remain focused on aggregate indicators; sectoral nowcasting (especially agriculture) is at an early stage with fewer applications compared to aggregate GDP nowcasting.
- Tourism studies largely focus on arrivals and overnight stays; comparatively fewer studies focus on revenue-based measures.
- Use of Google Trends and similar search-based indicators is well established in advanced economies for labor markets, with fewer applications in emerging markets.
- Taken together, literature suggests scope to extend non-traditional data use across more sectoral indicators and country settings.

### Satellite and Remote-Sensing Data for Agriculture Value Added
- Key technical properties and timeliness:
  - Sentinel-2: resolutions as fine as 10 meters across primary spectral bands.
  - MODIS vegetation index products: derived from daily atmosphere-corrected surface reflectance and composited into 16-day products.
  - Sentinel-2 constellation: revisit period of approximately five days.
- Widely used satellite-derived indicators and their monitoring roles:
  - Agriculture Stress Index (ASI): reports share of cropland affected by water stress; summarized monthly using latest dekadal (10-day) ASI value; used as early warning for drought and to flag recovery (example: elevated cropland stress in early 2025 and recovery in early 2026).
  - Precipitation: proxy for water availability; Figure 2 highlights contrast between drought in January 2025 and rainfall recovery in January 2026.
  - Normalized Difference Vegetation Index (NDVI): tracks vegetation health using red and near infra-red; shows high correlation with cereal production (Figure 3).
  - Enhanced Vegetation Index (EVI): improves on NDVI by reducing atmospheric/background effects and improving sensitivity in dense vegetation.
  - Vegetation Health Index (VHI): combines VCI and TCI to assess vegetation health and monitor drought.
  - NDVI Anomalies: deviations from historical average to identify unusual vegetation conditions.
  - Temperature: air temperature at 2 meters above surface to assess heat effects on cropland.
  - Nighttime lights: proxy for broader economic activity; regional tracking shows higher radiance in urban, industrial, logistics centers and an increase in luminosity between 2015 and 2025.
  - Groundwater Storage Percentile: current groundwater storage relative to historical distribution, expressed as a percentile.
- Exploratory modeling approach: include a broad set of indicators to capture diverse signals and subtle differences among related variables; correlations between these indicators and agricultural output vary across variables but provide relevant signals (Figure 5).

### Summary of Satellite-based Indicators: Frequency, Timeliness, and Sources (as reported)
- Agriculture Stress Index (ASI): Frequency = Dekadal (10 days); Timeliness = near real-time; Source = Agricultural stress index near real time (Global - Dekadal - 1 km ) - FAO ASIS
- Precipitation: Frequency = Daily; Timeliness = ~5-6 days; Source = ERA5-Land Daily Aggregated - ECMWF Climate Reanalysis
- NDVI (Normalized Difference Vegetation Index): Frequency = 8 days; Timeliness = ~1-2 weeks; Source = Landsat Collection 2 Tier 1 Level 2 8-Day NDVI Composite
- Enhanced Vegetation Index: Frequency = 8 days; Timeliness = ~1-2 weeks; Source = Landsat Collection 2 Tier 1 Level 2 8-Day EVI Composite
- Vegetation Health Index (VHI): Frequency = Dekadal (10 days); Timeliness = near real-time; Source = Vegetation health index near real time (Global - Dekadal - 1 km) - FAO ASIS
- NDVI Anomalies: Frequency = Dekadal (10 days); Timeliness = near real-time; Source = Non-Seasonal Indicators (Crops - Agricultural Stress Index System)
- Temperature: Frequency = Daily; Timeliness = ~5-6 days; Source = ERA5-Land Daily Aggregated - ECMWF Climate Reanalysis
- Nighttime Lights: Frequency = Daily; Timeliness = ~1-2 weeks; Source = VNP46A2: VIIRS Lunar Gap-Filled BRDF Nighttime Lights
- Groundwater Storage Percentile: Frequency = Weekly; Timeliness = ~2-9 days; Source = Groundwater storage percentile (GRACE-DA-DM Global v3.0, 0.25°, weekly)
- Note: Timeliness figures are approximate and reflect typical availability under normal processing conditions; actual availability may vary due to satellite downtime, processing lags, reprocessing cycles, or differences between preliminary and final validated product releases.

### High-frequency Proxies for Tourism Revenues
- FlightRadar24 / Passenger Capacity (IMF Statistics Department indicator):
  - Daily high-frequency indicator; available with a 5-day lag.
  - Designed to estimate total seating capacity on flights by estimating seat count from aircraft type.
  - Strong co-movement with official tourism arrivals: correlation of 0.9 in levels and 0.95 when smoothed using a 3-month moving average (Figure 6).
  - Data disaggregated by country and airport; enables detailed tracking of tourism activity (example: increased activity across major airports between December 2025 and January 2026 coinciding with AFCON 2025, Figure 7).
  - Official arrivals data: released with a 1-month lag; tourism revenues: published with a 1-month lag.
  - Passenger capacity is a granular and timely proxy for tourism arrivals, useful during peak events such as AFCON 2025 and the 2030 FIFA World Cup.
- Google Trends:
  - Reports relative search intensity scaled from zero to one hundred; available at daily, weekly, or monthly frequencies depending on query period and geographic scope.
  - Timeliness: near real-time availability at a daily frequency.
  - Established usefulness in literature for “predicting the present” and tracking destination-related searches for tourism demand.
  - Tourism revenues show notable associations with multiple Google Trends indices in different languages (Figures 8 and 9), supporting use as timely signals for nowcasting tourism revenues.
  - Note on normalization: Google Trends data are normalized and scaled from zero to one hundred; the index represents relative search interest rather than absolute search volumes.

### Google Trends and Unemployment Nowcasting
- Google Trends queries related to job hunting, employment platforms, and unemployment support provide timely signals of labor-market conditions.
- Literature evidence:
  - D’Amuri and Marcucci (2017): Google job-search intensity improves forecasting of U.S. unemployment rate.
  - Askitas and Zimmermann (2009): Internet search queries track and predict unemployment dynamics in Germany.
- Empirical associations in this work:
  - Unemployment rate closely associated with several Google Trends indices related to job search activity (Figures 10 and 11).
  - Correlation heatmaps and time-series comparisons reveal notable co-movement between search activity and labor market conditions.
  - Consistent with established evidence, findings reinforce potential use of Google Trends as timely inputs for nowcasting unemployment ahead of official release.

*Source: wpiea2026108-source-pdf - 2. Literature Review*

### 4. Methodology

### 4. Methodology

### 4.1. Data Transformation and Processing

- General
  - Target variables are growth rates or levels from official statistics; predictors are non-traditional, high-frequency indicators aggregated to the target frequency.
  - High-frequency indicators are aggregated to quarterly frequency using sum for flow variables and mean for indices, then expressed in year-on-year (YoY) growth rates to align with the target variable (where applicable).
  - For each satellite or non-traditional indicator, temporal vintages are constructed for contemporaneous quarter t and lags t-1 and t-2 (where specified) to capture information lags and carry-over effects.
  - Prior to estimation, highly correlated variables in the training set are filtered at a 90 percent correlation threshold to reduce redundancy.
  - Predictors are screened for data quality and stability; indices with excessive zeroes are filtered.

- 4.1.1. Agriculture Value Added
  - Target: year-on-year (YoY) growth rate of quarterly agricultural GVA in volumes computed from official national accounts data.
  - Predictors: satellite-derived vegetation and climate-related indicators available at daily, weekly, or dekadal (10-day) frequencies, aggregated to quarterly frequency (sum for flows; mean for indices) and expressed in YoY growth rates.
  - Temporal vintages: contemporaneous quarter t and two lags (t-1, t-2).
  - Pre-estimation filtering: highly correlated variables in the training set removed at a 90 percent threshold; predictors screened for data quality and stability.

- 4.1.2. Tourism Revenues
  - Target: monthly tourism receipts obtained from official statistics.
  - Predictors: monthly Google Trends indices and daily Passenger Capacity indicator derived from FlightRadar24 data, both aggregated to monthly frequency (mean for Google Trends indices; sum for Passenger Capacity).
  - Data extraction: Google Trends data were extracted once on April 3, 2026, using a fixed sample window and identical settings to ensure comparability of normalized (0–100) indices.
  - Transformations:
    - Passenger Capacity: log-difference transformation to account for seasonality and stationarity.
    - Google Trends: log(1+p) transformation applied, where p denotes the search interest index, to accommodate zero-valued observations.
  - Query set: a broad set of 100 tourism-related search terms in French, Spanish, and English covering accommodation platforms, flights, and travel planning.
  - Pre-estimation filtering: indices with excessive zeroes and highly correlated variables in the training set removed at a 90 percent threshold.
  - Lag evaluation: Google Trends indices evaluated from contemporaneous (t) to six months (t-6).

- 4.1.3. Unemployment Rate
  - Target: quarterly unemployment rate obtained from official statistics.
  - Predictors: monthly Google Trends search indices capturing unemployment-related search activity.
  - Data extraction: Google Trends data were extracted once on April 3, 2026, using a fixed sample window and identical settings to ensure comparability of normalized (0–100) indices.
  - Query set: a broad set of 100 search terms related to unemployment and job-seeking behavior in French, Arabic, and English covering vacancy listings, job-search platforms, unemployment benefits, and job application procedures.
  - Aggregation: Google Trends indices aggregated to quarterly frequency using mean aggregation to align with the target variable.
  - Pre-estimation filtering: indices with excessive zeroes and highly correlated variables in the training set removed at a 90 percent threshold.
  - Lag evaluation: alternative lag structures from contemporaneous (t) to two quarters (t-2) are evaluated.
  - Structural break: a dummy variable is included to capture the change in labor market dynamics since Q4 2019 in the context of the COVID pandemic.

### 4.2. Nowcasting Framework

- Overview
  - Twelve model specifications are estimated and compared across three families: linear models, tree-based ensemble methods, and time series models.
  - Rationale: different indicators and data involve different types, frequencies, and predictor-target relationships; combination of model families addresses linearity, non-linearities, interactions, and autoregressive dynamics.
  - To address high dimensionality, each model is paired with a variable selection method to retain the most informative predictors.
  - A univariate ARIMA model estimated on the target variable alone is used as the benchmark.

- 4.2.1. Models (specifications explored)
  - Random Forest
    - Ensemble of independently grown decision trees trained on bootstrap resamples and random subsets of features; predictions by averaging across trees.
    - Hyperparameters: number of trees, maximum tree depth, minimum samples per split and leaf node, size of feature subset at each split.
  - Random Forest with Variable Selection
    - Variable selection via Mean Decrease in Impurity (MDI) importance scores from a preliminary RF on training data.
    - Features with importance exceeding the mean MDI across candidates are retained; selection performed exclusively on training data to avoid lookahead bias.
  - XGBoost
    - Gradient-boosted tree ensemble adding trees sequentially to correct residuals; trees regularized by depth and shrinkage (learning rate).
  - XGBoost with Variable Selection
    - Variable selection via XGBoost gain-based importance; features with gain exceeding the mean gain across candidates are retained.
  - Elastic Net
    - Penalized linear regression with combined L1 (LASSO) and L2 (Ridge) penalties.
    - Hyperparameters: overall regularization strength α and mixing parameter ρ ∈ [0, 1] interpolating between Ridge (ρ = 0) and LASSO (ρ = 1).
  - Elastic Net with Variable Selection
    - Selection by estimating Elastic Net on full predictor set and retaining predictors with non-zero coefficients; model re-estimated on reduced set.
  - Partial Least Squares (PLS)
    - Dimension-reduction regression constructing latent components maximizing covariance with the target; main tuning parameter is number of retained components.
  - Partial Least Squares with Variable Selection
    - Selection via Variable Importance in Projection (VIP) scores; variables with VIP > 1 retained and model refit on reduced set.
  - Ordinary Least Squares (OLS)
    - Standard linear regression estimated by minimizing sum of squared residuals.
  - Ordinary Least Squares with Variable Selection
    - Variable selection via LASSO regularization; penalty parameter λ selected via k-fold cross-validation on training set; non-zero LASSO coefficients identify selected predictors, upon which OLS is estimated.
  - ARIMAX
    - ARIMA augmented with exogenous regressors characterized by p (autoregressive order), d (degree of differencing), q (moving average order).
  - ARIMAX with Variable Selection
    - Variables ranked by absolute marginal correlation with the target using Sure Independence Screening (SIS).
    - Retained variables standardized and reduced by Principal Component Analysis (PCA) fitted on training set; number of components chosen to explain 90 percent of the variance; components used as regressors in ARIMAX.
  - ARIMA
    - Pure time-series ARIMA on the target alone serving as primary benchmark.
    - Lag orders selected by exhaustive grid search over p, d, q minimizing Akaike Information Criterion (AIC); models estimated via maximum likelihood in a state-space representation with non-convergent specifications discarded.

- 4.2.2. Hyperparameter Tuning and Model Evaluation
  - Train-test split: dataset split into training set comprising 80 percent of the data and test set comprising the remaining 20 percent.
  - Hyperparameter selection: for tunable models (Random Forest, XGBoost, Elastic Net, PLS), selected within the training set using five-fold time-series cross-validation.
  - Model complexity rule: one-standard-error rule applied to retain the simplest configuration whose RMSE falls within one standard error of the minimum.
  - ARIMA and ARIMAX orders selected by minimizing AIC subject to convergence checks.
  - Performance metrics:
    - Primary: Root Mean Squared Error (RMSE).
    - Complementary: Mean Absolute Error (MAE).
    - Relative performance reported against the ARIMA benchmark.
  - Evaluation procedure: rolling pseudo-out-of-sample evaluation where at each test period the model is re-estimated using all available data up to that point and then used to generate the nowcast; estimation window expands over time to produce multiple forecast errors.

- Train and test coverage by indicator (Table 3)
  - Agricultural GVA
    - Training Dataset (80 percent): 2015Q1 - 2023Q3
    - Testing Dataset (20 percent): 2023Q4 - 2025Q4
    - Nowcast: 2026Q1
  - Tourism Revenues
    - Training Dataset (80 percent): 2019M11 - 2024M10
    - Testing Dataset (20 percent): 2024M11 - 2026M2
    - Nowcast: 2026M3
  - Unemployment Rate
    - Training Dataset (80 percent): 2006Q1 - 2021Q4
    - Testing Dataset (20 percent): 2022Q1 - 2025Q4
    - Nowcast: 2026Q1

### 4.2.3. Robustness Checks

- Three robustness checks conducted:
  - Sub-sample stability: the test period is divided into two equal sub-samples and model performance evaluated separately in each half to assess consistency in model ranking across sub-periods.
  - Iterative updates: the nowcasting framework is updated iteratively as new observations become available; the best-performing model in each sector retains its ranking across updates, providing real-time confirmation of stability.
  - Statistical significance: the Clark–West test is conducted on pseudo out-of-sample forecasts to assess whether the best-performing model provides a statistically significant improvement over the benchmark.

*Source: 4. Methodology, wpiea2026108-source-pdf*

### 5. Results

### 5. Results

### Overview
- All models described in section 4.2.1 are estimated for each variable and compared based on rolling pseudo-out-of-sample RMSE.
- For agricultural GVA nowcasting using satellite-based indices, Random Forest delivers the lowest out-of-sample error.
- Satellite-based indicators such as vegetation indices and nighttime lights exhibit complex nonlinear relationships that machine learning models are well-positioned to capture.
- For tourism revenues nowcasting using Google Trends indices and the FlightRadar24 passenger capacity, the best performing model is ARIMAX with PCA-based dimensionality reduction of the Google Trends terms and the passenger capacity as exogenous regressors.
- For unemployment rate nowcasting using Google Trends indices, the best performing model is ARIMAX with PCA-based dimensionality reduction of the Google Trends terms as exogenous regressors.
- For both tourism and unemployment, the ARIMAX specification with variable selection outperforms the specification that includes the full set of Google Trends predictors.
- The Clark–West test indicates statistically significant improvements at the 5 percent level over the benchmark across all three indicators, though the short evaluation windows mean this evidence should be interpreted with caution.
- Broad implication: model selection should be guided by the nature of the underlying data — machine learning for complex nonlinear satellite indicators, statistical time-series models for stable structured dynamics with exogenous inputs.

### 5.1. Agriculture Value Added
- Final predictor set (after initial pre-selection): EVI, NDVI anomalies, temperature, precipitation, groundwater storage percentile.
- Random Forest outperforms other statistical and machine learning models, delivering an 18 percent improvement in RMSE over the ARIMA benchmark.
- Random Forest has the lowest RMSE and MAE, indicating strong performance on average errors and sensitivity to larger deviations.
- The model captures key turning points, including the rebound in agriculture in early 2026; Q1/26 nowcasts increase with each new data release as the quarter advances, reflecting improved agricultural conditions ahead of official data releases expected in end-June 2026.
- Shapley decomposition for the 2026Q1 nowcast: nowcast is primarily driven by strong vegetation signals across several months, followed by temperature and groundwater storage percentile (Shapley contributions indicate relative feature importance but do not imply causality).

- Table 4. Agriculture GVA Nowcasting Model Performance (Model — Model Family — RMSE — MAE)
  - RF — Tree-based — 2.82 — 2.35
  - ARIMA — Time series — 3.46 — 2.68
  - ARIMAX_VS — Time series — 5.06 — 3.41
  - RF_VS — Tree-based — 6.07 — 4.33
  - ARIMAX — Time series — 6.20 — 5.46
  - XGB — Tree-based — 6.94 — 6.44
  - XGB_VS — Tree-based — 6.94 — 6.44
  - EN — Linear — 8.03 — 5.18
  - PLS_VS — Linear — 8.08 — 6.51
  - PLS — Linear — 9.20 — 7.68
  - EN_VS — Linear — 9.26 — 6.93
  - OLS — Linear — 12.39 — 10.77
  - OLS_VS — Linear — 12.39 — 10.77

### 5.2. Tourism Revenues
- Contemporaneous Google Trends signals outperform lagged specifications, suggesting within-month searches capture information realized in the same reporting period.
- Best performing model: ARIMAX with PCA-based dimensionality reduction, retaining two principal components from pre-selected Google Trends series, with passenger capacity from FlightRadar24 as a direct exogenous regressor.
- Model achieves a 14 percent improvement in RMSE relative to the ARIMA baseline.
- The model has the lowest RMSE and MAE across benchmarks and captures the overall impact of the 2025 Africa Cup of Nations (AFCON) event, with slight smoothing across December 2025 and January 2026.
- Estimated coefficients in the ARIMAX: passenger capacity coefficient = 0.5 (largest coefficient, dominant driver); Google Trends principal components contribute modestly but add incremental predictive power beyond flight capacity alone.

- Table 5. Tourism Revenues Nowcasting Model Performance (Model — Model Family — RMSE — MAE)
  - ARIMAX_VS — Time series — 0.19 — 0.14
  - ARIMA — Time series — 0.22 — 0.16
  - RF — Tree-based — 0.22 — 0.17
  - RF_VS — Tree-based — 0.23 — 0.18
  - XGB — Tree-based — 0.28 — 0.21
  - XGB_VS — Tree-based — 0.28 — 0.21
  - PLS — Linear — 0.32 — 0.25
  - PLS_VS — Linear — 0.43 — 0.35
  - EN_VS — Linear — 0.46 — 0.38
  - EN — Linear — 0.48 — 0.39
  - OLS_VS — Linear — 0.55 — 0.44
  - OLS — Linear — 2.53 — 1.44
  - ARIMAX — Time series — 2.73 — 2.04

### 5.3. Unemployment Rate
- Contemporaneous Google Trends signals outperform lagged specifications, reflecting real-time job-related behavior.
- Best performing model: ARIMAX with PCA-based dimensionality reduction retaining six principal components from pre-selected Google Trends series, plus a dummy variable capturing the structural break following COVID as exogenous regressors.
- Model achieves a 25 percent improvement in RMSE relative to the ARIMA baseline.
- The best-performing model records the lowest values on both RMSE and MAE and captures the downward trend in unemployment following the end of the drought period.
- Principal component analysis aggregates diverse search behavior (terms across fr, ar, en) into components related to job applications, resignations, internships, and unemployment; loadings indicate contribution direction and magnitude but are not causal.

- Table 6. Unemployment Rate Nowcasting Model Performance (Model — Model Family — RMSE — MAE)
  - ARIMAX_VS — Time series — 0.44 — 0.31
  - ARIMA — Time series — 0.59 — 0.45
  - RF_VS — Tree-based — 0.69 — 0.55
  - RF — Tree-based — 0.72 — 0.58
  - OLS_VS — Linear — 0.79 — 0.68
  - PLS_VS — Linear — 0.85 — 0.74
  - XGB — Tree-based — 0.92 — 0.74
  - XGB_VS — Tree-based — 0.92 — 0.74
  - PLS — Linear — 1.05 — 0.95
  - EN — Linear — 1.64 — 1.50
  - EN_VS — Linear — 1.73 — 1.60
  - OLS — Linear — 1.99 — 1.51
  - ARIMAX — Time series — 2.20 — 1.86

### Conclusions (from Section 6, summarized here)
- Non-traditional data sources can meaningfully complement official statistics to produce more timely and more accurate nowcasts of key macroeconomic indicators for Morocco.
- Across agriculture, tourism, and unemployment, models incorporating non-traditional data consistently surpass benchmarks based only on standard time-series techniques.
- No single method outperforms across all indicators; performance advantage of machine learning versus traditional statistical models varies by sector and data type.
- The data sources and nowcasting framework are explicitly designed for replicability in other countries and can be adapted to sectors most relevant to each country (e.g., industry, construction, retail).
- Future work: explore additional non-traditional data types, investigate alternative machine learning and statistical model configurations, and leverage the pipeline across countries with scarce data to provide policymakers more timely estimates of key economic indicators.

*Source: IMF Working Paper No. WP/2026/108, Section 5 and concluding remarks.*

---


_Source: https://www.imf.org/-/media/files/publications/wp/2026/english/wpiea2026108-source-pdf.pdf_