## Introduction

## Source details

**Canonical URL:** [Introduction](https://www.imf.org/-/media/files/publications/wp/2025/english/wpiea2025222-source-pdf.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/wp/2025/english/wpiea2025222-source-pdf.pdf.md)
- [Structured JSON version](/-/media/files/publications/wp/2025/english/wpiea2025222-source-pdf.pdf.json)

---

### Context and motivation
- Governments increasingly use industrial policies—selective government interventions targeting specific firms, sectors, or geographic regions—because private sector initiative alone will not attain societal goals.
- Contemporary industrial policy differs from post-1945 state ownership and control; motives have widened and large public funds are being allocated, raising public finance concerns.
- There is no global agency systematically collecting information on industrial policy interventions and no consensus on which interventions constitute industrial policy; the intertemporal dimension of policy choice is underdeveloped.
- The risk of policy mistakes is high, and in an interconnected trading system with limited agreement on appropriate industrial policy choices, the risk of escalating trade disputes is ever present.

### New dataset: Historical NIPO (H-NIPO)
- Constructed as a historical extension of the New Industrial Policy Observatory (NIPO) covering 2009-2023 using Global Trade Alert (GTA) records that meet the IMF definition of industrial policy.
- Inclusion criteria:
  - A GTA measure is included if it meets either (1) an inferred industrial policy motive (one of four motives) identified via an LLM trained on NIPO 2023 motives, or (2) targets certain strategic sectors based on pre-defined product lists used in NIPO.
- Key dataset statistics and composition:
  - Encompasses 34,248 distinct policy interventions worldwide with announcement, implementation, and possible withdrawal dates between 2009-2023.
  - 85 percent of interventions are trade distortive; 15 percent are liberalizing on a non-discriminatory basis.
  - 52 percent of measures are by Advanced Economies (AEs); 48 percent by Emerging Market and Developing Economies (EMDEs).
  - China, the European Union, and the United States together account for around 53 percent of interventions.
  - Coverage: state measures implemented or announced between January 1, 2009 and December 31, 2023.
  - Monitoring sample: 75 jurisdictions making up 94 percent of global GDP.
  - Jurisdiction composition: 53.6 percent AEs and 46.4 percent EMDEs.
  - Regional breakdown examples: 13 jurisdictions in Asia Pacific (28.5%), 31 in Europe and Central Asia (28.6%), 9 in Latin American and the Caribbean (11.5%), 9 in Middle East and North Africa (1.7%), 2 in North America (21.3%), 4 in South Asia (6.8%), 7 in Sub-Saharan Africa (1.6%).

### Stylized facts from 2009–2023
- Structural shift:
  - 2020 marks a turning point: enhanced resort to selective policy intervention; the number of new interventions did not diminish as countries exited the COVID-19 pandemic.
  - November 2019 identified as the most statistically significant breakpoint marking an acceleration of policy activity (supremum Wald test over 2009-2023).
- Motives:
  - During and after the GFC the most frequently cited motives were competitiveness and climate.
  - Since 2020, motives such as supply chain resilience, national security, and geopolitical concerns have become prominent.
- Instruments and cross-country patterns:
  - Subsidies were consistently the main instrument in both AEs and EMDEs over 2009-2023.
  - Trade measures were more commonly deployed by EMDEs.
  - Convergence in instrument use between AEs and EMDEs between 2009-2023:
    - Share of subsidies in AEs decreased from 84 to 75 percent.
    - Share of subsidies in EMDEs rose from 56 to 71 percent.
    - Share of trade measures in AEs increased from 3 to 8 percent.
    - Share of trade measures in EMDEs decreased from 27 to 18 percent.
  - Public procurement localization measures account for over 5 percent of trade-distorting industrial policies in AEs.
  - Import restricting measures account for 18 percent of all trade distorting industrial policies in EMDEs.
- Green industrial policies:
  - Large share of green industrial policies is a more recent phenomenon linked to green stimulus after the 2007-2008 GFC and the COVID-19 pandemic.
  - Examples in the source: 2008 European Recovery Plan; 2020 European Green Deal and Next Generation EU package; American Recovery and Reinvestment Act of 2009 (with an estimated around 17% of all direct spending on clean energy); Inflation Reduction Act of 2022 tax credits; China’s post-COVID-19 green stimulus focused on high-speed rail, electricity networks and water management.

### Drivers of post-2020 changes (empirical findings)
- Key sign reversals and shifts around 2020:
  - Exposure to imports from geopolitically distant partners shifted from being negatively correlated to positively correlated with industrial policies targeting that product—consistent with derisking calls gaining traction.
  - Industrial support by other countries for a product turned from a negative to a positive driver of policy intervention—consistent with renewed tit-for-tat dynamics.
  - Correlation with comparative advantage switched from positive to negative—implying a shift to newer, less established sectors.
  - Correlation with the stock of existing industrial policies targeting a product changed from negative to positive—consistent with acceleration and growing concentration of support for favored sectors.
- Heterogeneity by country income and motive:
  - The post-2020 shift toward policies targeting products with higher geopolitical exposure is driven by EMDEs.
  - The increased concentration of industrial policy measures in sectors with a higher stock of existing support is driven by AEs.
  - Measures motivated by national security and supply chain resilience play a prominent role in accounting for observed shifts, including derisking geopolitical exposure, tit-for-tat dynamics, and increased acceleration and concentration of support.

### Methodology summary for H-NIPO construction
- Inclusion rules:
  - A GTA measure is included if it is associated with at least one of four inferred motives or if it targets at least one of a pre-defined set of product or service categories (HS and CPC codes).
- Motives included (LLM-inferred):
  - National security and/or geopolitical concerns (treated as a single motive in H-NIPO).
  - Resilience/security of supply chains (non-food).
  - Domestic competitiveness in strategic sectors.
  - Climate change mitigation and other environmental objectives.
- Product and service inclusion categories (six-digit HS subheadings and CPC codes):
  - Low-Carbon Technology.
  - Dual-use products.
  - Critical minerals.
  - Advanced technology products.
  - Medical products.
  - Chemicals.
  - Critical Raw Materials Downstream Industry.
  - Industrial Raw Materials.
  - IT or digital services (identified by CPC codes 623, 831, 834, 839, 841, 842, 843).
- Broad policy instrument taxonomy (58 specific instruments grouped into 7 types, examples):
  - Export barriers (export bans, tariffs, quotas, export licensing, other export-related trade barriers).
  - Import barriers (import bans, tariffs, quotas, import licensing, other import-related trade barriers).
  - Domestic subsidies (tax rebates, grants, state loans, loan guarantees, price stabilization, production subsidies, other incentives).
  - Export incentives (tax-based export incentives, unit-based export subsidies, trade financing, other financial export promotion).
  - Foreign Direct Investment measures (entry and ownership requirements, FDI screening decisions).
  - Procurement policies (changes to public procurement law or practice that may favor local suppliers).
  - Localization incentives or requirements (including public procurement localization measures).
- Other variables recorded:
  - Intervention ID and Title; Jurisdiction; Level of government implementation; Initial assessment on relative treatment (distortive or liberalizing); Announcement, implementation, and removal date; Targeted economic activity (HS 6-digit and CPC 3-digit where available); Source.
  - H-NIPO does not record monetary values associated with subsidies (GTA only systematically collects such values since 2023).

### LLM-based motive assignment (Annex III summary)
- Model and training:
  - Model: RoBERTa with a sequence classification layer; full finetuning of all model weights.
  - Training: class-stratified 80% training / 20% testing split; repeated for different seeds; trained for 10 epochs.
  - Optimizer: AdamW; learning rate: 1×10^−5; weight decay: 0.01; linear learning rate scheduler with warm-up proportion 10%.
- Ensemble prediction procedure:
  - Train RoBERTa ten times with randomized classification head initialization and randomized batch order.
  - For each policy, run the ten finetuned models to obtain ten sets of probabilities per motive.
  - Compute weights for each set using the F1 score from the testing dataset and compute a weighted average probability.
  - Classify a policy as having a particular inferred motive if the weighted probability exceeded 60%.
  - A policy can be associated with between 0 and 4 motives.
- Performance (average precision and recall across the 10 fine-tuned models):
  - Without motive:
    - precision 0.95
    - recall 0.93
  - Strategic competitiveness:
    - precision 0.80
    - recall 0.83
  - GVC Resilience:
    - precision 0.83
    - recall 0.75
  - Geopolitical concerns or national security:
    - precision 0.90
    - recall 0.93
  - Climate change mitigation:
    - precision 0.84
    - recall 0.87
  - Summary: algorithm performs best for policies without motive, achieves a precision above 80% for all motives and recall above 80% for three out of four motives.

### Structural trends in trade-distorting industrial policy (2009-2023)
- Emphasis on cumulative number of implemented distortive policies (85 percent of policies recorded).
- Target sectors and trends:
  - Majority of trade-distorting measures target products from predefined lists of strategic sectors: military/civilian dual use products, advanced technologies, and upstream inputs such as critical minerals, chemicals, steel, and aluminum.
  - Low carbon technology products do not account for a prominent share despite climate mitigation motives; many climate-motivated measures target energy intensive industries (steel and aluminum).
  - Policies targeting critical minerals and industrial raw materials have become more salient in recent years.
- Instruments by income group and changes over time (reiterated):
  - Financial grants most common among EMDEs; state loans and trade finance dominate subsidies in AEs.
  - Convergence between AEs and EMDEs (2009-2023):
    - Share of subsidies in AEs decreased from 84 to 75 percent.
    - Share of subsidies in EMDEs rose from 56 to 71 percent.
    - Share of trade measures in AEs increased from 3 to 8 percent.
    - Share of trade measures in EMDEs decreased from 27 to 18 percent.
    - Localization policies increased from 7 to 8 percent in AEs and from 2 to 6 percent in EMDEs.
- Instrument mixes by motive (2009-2019 vs 2020-2023; Table A1 highlights):
  - Climate Change Mitigation:
    - Domestic subsidy: 89; 92
    - Localization or Procurement: 5; 5
  - National Security and/or Geopolitical Concern:
    - Domestic subsidy: 6; 10
    - Export barrier: 7; 22
    - FDI: 24; 11
    - Import barrier: 28; 11
    - Localization or Procurement: 30; 9
    - Other: 5; 37
  - Resilience/Security of Supply Chains (Non-food):
    - Domestic subsidy: 82; 76
    - Export barrier: 3; 13
  - Strategic Competitiveness:
    - Domestic subsidy: 63; 75
    - Export subsidy: 27; 11
    - Import barrier: 4; 7

### Empirical approach and key findings (country-product level)
- Dependent variable:
  - Change in implemented trade distortive industrial policy measures at the country c × product p (six-digit HS subheading) level.
- Benchmark specification:
  - ∆IPMeasures_{c,t,p} = α_{ct} + γ_{pt} + θ_{cp} + β X_{p,c,t-1} + ε_{c,t,p}
- Key explanatory variables in X_{p,c,t-1} include:
  - Import-weighted geopolitical distance (GPD).
  - Export-weighted industrial policy measures by other countries (XWIP).
  - Herfindahl–Hirschman index (HHI) of a country’s import concentration for product p.
  - Change and level of Balassa revealed comparative advantage (RCA).
  - Count/stock of past industrial policies (Count of IPs).
- Main empirical findings (rolling six-year subperiods):
  - Structural break around 2020 with several key coefficients exhibiting sign reversals.
  - Geopolitical exposure (GPD): negative until 2012-2018, became insignificant, and turned positive and significant in recent years (example coefficient path includes -0.0588***, …, 0.0433***, 0.0298**).
  - Export exposure to other countries’ industrial policies (XWIP): negative until about 2014-2020 / 2015-2021, then flipped to positive and significant afterwards (example path -0.3109***, …, 0.4197***, 0.4523***).
  - Import concentration (HHI): consistently positive and significant (example path 0.0688***, …, 0.0950***, 0.0879***).
  - RCA: level becomes significant in recent periods with a sign reversal between 2014-2020 and 2016-2022.
  - Count of IPs: sign reversal from negative to positive between 2012-2018 and 2014-2020.
- Heterogeneity by income group:
  - Aggregate shifts mask differences: EMDEs drive the post-2020 shift toward policies targeting products with higher geopolitical exposure; AEs drive increased concentration on sectors with a higher stock of existing support.
  - For EMDEs the Count of IPs coefficient is consistently negative and significant in the second period; for AEs the coefficient turns from negative and significant to positive and significant.

### Limitations and caveats
- GTA does not include regulatory standards, TBTs, or SPS measures; industrial policy pursued through product market regulation and standards is not captured.
- H-NIPO is a subset of the GTA database and relies on GTA inclusion rules; recording lags may exist.
- H-NIPO treats national security and geopolitical concerns as a single motive and does not include "digital transformation" as a motive (digital transformation was added later in NIPO).
- H-NIPO does not record subsidy monetary values systematically (GTA began systematic collection of subsidy values in 2023).

### Conclusions and research implications
- 2020 marks a structural shift toward increased use of industrial policy in both AEs and EMDEs.
- Subsidies are the main policy instrument in both AEs and EMDEs; the share of countries implementing them has risen markedly since the GFC.
- Empirical evidence is consistent with derisking (geopolitical exposure), tit-for-tat dynamics (export exposure to other countries’ policies), and increasing sectoral concentration (stock of past policies).
- Revealed comparative advantage switched from a positive to a negative correlation (driven by AEs), indicating a shift toward less established sectors.
- NS/GVC-motivated measures account for many observed shifts in behavior.
- Dataset contribution and suggested future research:
  - H-NIPO provides a long-running inventory for studying dynamic impacts of non-market interventions on comparative advantage, motive-specific effectiveness, welfare implications, and international spillovers relevant to WTO reform debates.
  - Suggested future work: exploit the over 15-year horizon to assess dynamic impacts, motive-specific success, net welfare effects, and cross-country spillovers.

*Source — Introduction, wpiea2025222-source-pdf (Historical NIPO methodology and findings).*

### Introduction ...........................................................................................................

### Introduction

### Covered Sections
- Introduction ......................................................................................................................................................... 5
- Methodology of the Historical NIPO database ................................................................................................. 9
- Inclusion in the H-NIPO database ................................................................................................................ 10
- Motives .................................................................................................................................................. 10
- Assigning Motives using Large Language Model .................................................................................. 11
- Product or Service Categories ............................................................................................................... 11
- Broad Categories of Policy Instruments ................................................................................................ 12
- Other Variables in the H-NIPO Database .............................................................................................. 15
- Structural Trends in Global Industrial Policy: 2009-2023 .............................................................................. 16
- Determinants of Industrial Policy Use ............................................................................................................ 21
- Conclusions ....................................................................................................................................................... 26

### Annexes and Methodological Appendices
- Annex I. Methodology for Global Trade Alert Data Collection ...................................................................... 27
- Annex II. List of Jurisdictions in the Extended NIPO Database ................................................................... 28
- Annex III. Classifying Industrial Policy Motives using a Large Language Model Ensemble Approach ... 29
- Annex IV. Additional Descriptive Statistics .................................................................................................... 31
- References ......................................................................................................................................................... 32

### Figures (listed)
- Figure 1. 2020 marks a turning point in industrial policy activity ............................................................................ 16
- Figure 2. New industrial policies with LLM-assigned motive ................................................................................... 17
- Figure 3. New industrial policies by sector ................................................................................................................ 18
- Figure 4. Trade distortive industrial policy instruments by income group ............................................................. 18
- Figure 5. Relationship between tax revenue and industrial subsidies .................................................................... 19
- Figure 6. The share of jurisdictions employing trade distortive policies has increased ....................................... 20

### Tables (listed)
- Table 1. Taxonomy of Specific Policy Institutions ................................................................................................... 14
- Table 2. The Evolving Relevance of Correlates of IPs ............................................................................................. 23
- Table 3. Differences in the Correlates of IP changes, Across Income Groups ..................................................... 24
- Table 4. Correlates of IP Dynamics, by Stated Motive ............................................................................................. 25
- A1. Frequency of Policy Instrument by Stated Motive .................................................................................. 31

### Key methodological highlights (as described in section titles)
- The document includes a "Methodology of the Historical NIPO database" (page 9).
- Inclusion criteria are discussed under "Inclusion in the H-NIPO database" (page 10).
- Motive classification is addressed in "Motives" (page 10) and in detail under "Assigning Motives using Large Language Model" (page 11).
- Product or service categorization is described in "Product or Service Categories" (page 11).
- Policy instrument classification is organized under "Broad Categories of Policy Instruments" (page 12).
- Additional dataset variables are reported in "Other Variables in the H-NIPO Database" (page 15).
- Annex III elaborates on "Classifying Industrial Policy Motives using a Large Language Model Ensemble Approach" (page 29).
- Annex I documents the "Methodology for Global Trade Alert Data Collection" (page 27).

### Thematic focus areas (inferred from section and figure/table headings)
- Historical compilation and methodological transparency for the H-NIPO database (sections on methodology, inclusion, variables).
- Use of LLMs for motive assignment and classification (section and Annex III).
- Cross-cutting structural trends in global industrial policy during 2009-2023 (section and Figures 1–3).
- Analysis of trade-distorting instruments and their distribution by income group (Figure 4; Figure 6).
- Links between fiscal capacity (tax revenue) and industrial subsidies (Figure 5).
- Empirical analysis of determinants and correlates of industrial policy adoption and dynamics (sections and Tables 2–4).
- Supplementary descriptive statistics and jurisdiction lists in Annexes II and IV.

_This overlay summarizes the structure, methodological components, figures, and tables listed in the Introduction section of the source PDF._

### Introduction

### Introduction

### Context and motivation
- Governments have increasingly turned to industrial policies—defined here as selective government interventions targeting specific firms, sectors, or geographic regions—because private sector initiative alone will not attain societal goals.
- Contemporary industrial policy differs from post-1945 state ownership and control; motives have widened and large public funds are being allocated, raising public finance concerns.
- There is no global agency systematically collecting information on industrial policy interventions and no consensus on which interventions constitute industrial policy; the intertemporal dimension of policy choice is underdeveloped.
- The risk of policy mistakes is high, and in an interconnected trading system with limited agreement on appropriate industrial policy choices, the risk of escalating trade disputes is ever present.

### New dataset: Historical NIPO (H-NIPO)
- H-NIPO is a historical extension of the New Industrial Policy Observatory (NIPO) covering 2009-2023 and is constructed from Global Trade Alert (GTA) records that meet the IMF definition of industrial policy (Industrial Policy Coverage in IMF Surveillance—Broad Considerations).
- Inclusion criteria: a GTA measure is included if it meets either (1) an inferred industrial policy motive (one of four motives) identified via an LLM trained on NIPO 2023 motives, or (2) targets certain strategic sectors based on pre-defined product lists used in the NIPO.
- The resulting H-NIPO dataset:
  - Encompasses 34,248 distinct policy interventions worldwide with announcement, implementation, and possible withdrawal dates between 2009-2023.
  - 85 percent of interventions are trade distortive; 15 percent are liberalizing on a non-discriminatory basis.
  - 52 percent of measures are by Advanced Economies (AEs); 48 percent by Emerging Market and Developing Economies (EMDEs).
  - China, the European Union, and the United States together account for around 53 percent of interventions.

### Stylized facts from 2009–2023
- Structural shift in 2020: enhanced resort to selective policy intervention; the number of new interventions did not diminish as countries exited the COVID-19 pandemic.
- Motives:
  - During and after the GFC the most frequently cited motives were competitiveness and climate.
  - Since 2020, motives such as supply chain resilience, national security, and geopolitical concerns have become prominent.
- Instruments:
  - Subsidies were consistently the main instrument in both AEs and EMDEs over 2009-2023.
  - Trade measures were more commonly deployed by EMDEs.
  - Convergence in instrument use between AEs and EMDEs between 2009-2023:
    - Share of subsidies in AEs decreased from 84 to 75 percent.
    - Share of subsidies in EMDEs rose from 56 to 71 percent.
    - Share of trade measures in AEs increased from 3 to 8 percent.
    - Share of trade measures in EMDEs decreased from 27 to 18 percent.
- Rise of green industrial policies:
  - Large share of green industrial policies is a more recent phenomenon linked to green stimulus after the 2007-2008 GFC and the COVID-19 pandemic.
  - Examples cited in the period include the 2008 European Recovery Plan, the 2020 European Green Deal and Next Generation EU package, the American Recovery and Reinvestment Act of 2009 (with an estimated around 17% of all direct spending on clean energy), and the Inflation Reduction Act of 2022 tax credits.
  - China implemented green stimulus post-COVID-19 focusing on high-speed rail, electricity networks and water management.

### Drivers of post-2020 changes (empirical findings)
- Around 2020 notable shifts in determinants of industrial policy targeting occurred:
  - Exposure to imports from geopolitically distant partners shifted from being negatively correlated to positively correlated with industrial policies targeting that product—consistent with derisking calls gaining traction.
  - Industrial support by other countries for a product turned from a negative to a positive driver of policy intervention—consistent with renewed tit-for-tat dynamics.
  - Correlation with comparative advantage switched from positive to negative—implying a shift to newer, less established sectors.
  - Correlation with the stock of existing industrial policies targeting a product changed from negative to positive—consistent with acceleration and growing concentration of support for favored sectors.
- Heterogeneity by country income and motive:
  - The post-2020 shift toward policies targeting products with higher geopolitical exposure is driven by EMDEs.
  - The increased concentration of industrial policy measures in sectors with a higher stock of existing support is driven by AEs.
  - Measures motivated by national security and supply chain resilience play a prominent role in accounting for observed shifts, including derisking geopolitical exposure, tit-for-tat dynamics, and increased acceleration and concentration of support.

### Related literature and data approaches
- Principal approaches to documenting industrial policy:
  - Inventories of policy interventions documented consistently over time (e.g., GTA-based approaches, including Juhász et al. (2023)).
    - GTA database includes records of over 77,700 policy interventions since 1 November 2008, covering subnational, central, and supranational levels.
  - Declarations by industrial groups or inferences from corporate financial statements (e.g., OECD MAGIC database Version 2.0 with subsidy receipt information for 482 industrial groups in 14 sectors from 2005-2022).
  - Applications of LLMs to published state documents across government levels (e.g., Fang, Li and Lu (2025) on China 2000-2022; Ju, Li, and Wei (2025) on US 1973-2022).
- Distinctions across datasets:
  - GTA-based approaches focus on state measures recorded by GTA and may differ from firm-level subsidy disclosures in MAGIC or from regulatory measures included by some LLM-based studies.
  - Methodological choices affect country and instrument representation (e.g., Juhász et al. (2023) find top users are AEs and exclude firm-level subsidy disclosures that make up much of China’s GTA-listed subsidies).

### Methodology summary for H-NIPO construction
- Coverage:
  - H-NIPO contains state measures implemented or announced between January 1, 2009 and December 31, 2023.
  - The monitoring covers a sample of 75 jurisdictions making up 94 percent of global GDP (selection based on consistent GTA tracking and G-20 core).
  - Jurisdiction composition: 53.6 percent AEs and 46.4 percent EMDEs.
  - Regional breakdown examples: 13 jurisdictions in Asia Pacific (28.5%), 31 in Europe and Central Asia (28.6%), 9 in Latin American and the Caribbean (11.5%), 9 in Middle East and North Africa (1.7%), 2 in North America (21.3%), 4 in South Asia (6.8%), 7 in Sub-Saharan Africa (1.6%).
- Inclusion rules:
  - A GTA measure is included if it is associated with at least one of four inferred motives or if it targets at least one of a pre-defined set of product or service categories (HS and CPC codes).
- Motives included (LLM-inferred):
  - National security and/or geopolitical concerns (treated as a single motive in H-NIPO).
  - Resilience/security of supply chains (non-food).
  - Domestic competitiveness in strategic sectors.
  - Climate change mitigation and other environmental objectives.
- LLM-based motive assignment:
  - Model: RoBERTa with a sequence classification layer.
  - Training: fine-tuned on the NIPO 2023 dataset; split into training and testing; trained for 10 epochs.
  - Prediction procedure:
    - Run model ten times to obtain ten sets of probabilities per intervention per motive.
    - Compute weights for each set using the F1 score from the testing dataset.
    - Compute weighted average probability.
    - Classify a policy as having a particular inferred motive if the weighted probability exceeded 60%.
    - Each policy can be associated with between 0 and 4 motives.
- Product and service inclusion categories (defined by six-digit HS subheadings and CPC codes):
  - Low-Carbon Technology.
  - Dual-use products.
  - Critical minerals.
  - Advanced technology products.
  - Medical products.
  - Chemicals.
  - Critical Raw Materials Downstream Industry.
  - Industrial Raw Materials.
  - IT or digital services (identified by CPC codes 623, 831, 834, 839, 841, 842, 843).
- Broad policy instrument categories covered (dataset includes 58 specific instruments grouped into 7 types):
  - Export barriers (export bans, tariffs, quotas, export licensing, other export-related trade barriers).
  - Import barriers (import bans, tariffs, quotas, import licensing, other import-related trade barriers).
  - Domestic subsidies (tax rebates, grants, state loans, loan guarantees, price stabilization, production subsidies, other incentives).
  - Export incentives (tax-based export incentives, unit-based export subsidies, trade financing, other financial export promotion).
  - Foreign Direct Investment measures (entry and ownership requirements, FDI screening decisions).
  - Procurement policies (changes to public procurement law or practice that may favor local suppliers).
  - Localization incentives or requirements (including public procurement localization measures).

### Limitations and caveats
- The GTA database does not include regulatory standards, TBTs, or SPS measures, so the dataset does not capture industrial policy pursued through product market regulation and standards.
- H-NIPO is a subset of the GTA database and relies on GTA inclusion rules; recording lags may exist.
- The H-NIPO treats national security and geopolitical concerns as a single motive and does not include "digital transformation" as a motive (digital transformation was added later in NIPO).

_Italic: Source — Introduction, wpiea2025222-source-pdf (Historical NIPO methodology and findings)._

### 1. Export ban

### 1. Export ban

### Policy taxonomy and related instruments
- The source lists export measures tracked in the GTA/NIPO taxonomy, including:
  - 1. Export ban
  - 2. Export licensing requirement
  - 3. Export quota
  - 4. Export tariff quota
  - 5. Export tax
  - 6. Local supply requirement for exports
  - 7. Export-related non-tariff measure, nes
- Import barriers enumerated in the taxonomy:
  - 1. Anti-dumping
  - 2. Anti-subsidy
  - 3. Import ban
  - 4. Import licensing requirement
  - 5. Import monitoring
  - 6. Import quota
  - 7. Import tariff
  - 8. Import tariff quota
  - 9. Internal taxation of imports
  - 10. Import-related non-tariff measure, nes
- Domestic subsidies instruments listed:
  - 1. Capital injection and equity stakes (including bailouts)
  - 2. Financial grant
  - 3. In-kind grant
  - 4. Tax or social insurance relief
  - 5. Production subsidy
  - 6. Interest payment subsidy
  - 7. Loan guarantee
  - 8. Import incentive
  - 9. Price stabilization
  - 10. State loan
  - 11. State aid, nes
  - 12. State aid, unspecified
- Export incentives:
  - 1. Trade finance
  - 2. Export subsidy
  - 3. Tax-based export incentive
  - 4. Financial assistance in foreign market
  - 5. Other export incentive
- FDI measures:
  - 1. FDI: Entry and ownership rule
  - 2. FDI: Financial incentive
  - 3. FDI: Treatment and operations, nes
- Public procurement measures:
  - 1. Public procurement access
  - 2. Public procurement, nes
- Localization content measures:
  - 1. Local content incentive
  - 2. Local content requirement
  - 3. Local operations incentive
  - 4. Local operations requirement
  - 5. Local value-added incentive
  - 6. Public procurement localization
  - 7. Localization, nes
- Other tracked instruments:
  - 1. Anti-circumvention
  - 2. Control on personal transactions
  - 3. Controls on commercial transactions and investment instruments
  - 4. Controls on credit operations
  - 5. Foreign customer limit
  - 6. Intellectual property protection
  - 7. Labor market access
  - 8. Post-migration treatment
  - 9. Repatriation & surrender requirements
  - 10. Special safeguard
  - 11. Trade payment measure
  - 12. Instrument unclear
- Note: “nes” refers to “not elsewhere specified” and in this context “elsewhere” refers to the policy interventions specifically named as those tracked in the GTA database.

### Other variables recorded in the H-NIPO database
- Intervention ID and Title: unique ID and title as in the GTA database.
- Jurisdiction: jurisdiction implementing the policy intervention or proposal.
- Level of government implementation: differentiates supra-national, national, and sub-national announcements.
- Initial assessment on relative treatment of domestic and foreign commercial interests: classifies measures as distortive or liberalizing based on likely effects on market competition and discrimination against foreign commercial interests.
- Announcement, implementation, and removal date: issuance, entry-into-force, and withdrawal/replacement dates; NIPO only includes measures announced or implemented from 1 January 2023.
- Targeted economic activity: affected HS codes at the 6-digit level (UN Harmonized System version 2012) and CPC sector codes at the 3-digit level (UN CPC codes, version 2.1) where available; sectors selected using UN correspondence table for CPC 2.1 and HS 2012 or GTA wording when product-level information absent.
- Source: the sources documenting the policy intervention or proposal.
- The H-NIPO database does not record values associated with subsidies, since this information has only been systematically collected by the GTA since 2023.

### Structural trends in trade-distorting industrial policy (2009-2023)
- Focus and scope:
  - Emphasis on the cumulative number of implemented distortive policies, which make up 85 percent of policies recorded in the database.
  - Study period and analysis span 2009-2023.
- Breakpoint and timing:
  - November 2019 identified as the most statistically significant breakpoint marking an acceleration of policy activity.
  - The breakpoint is determined by a supremum Wald test evaluating feasible breakpoints over 2009-2023 subject to minimum-sample constraints.
  - Similar breakpoints emerge around 2020 across income groups, motives, and instrument categories.
  - Chinese firm-level data exhibit excessive volatility and are excluded from the breakpoint analysis.
- Motives over time:
  - Strategic competitiveness and climate change mitigation were predominant motives following the GFC.
  - National security, geopolitical concerns, and supply chain resilience have increased in importance in more recent years, particularly post-pandemic.
  - For measures with multiple motives, each motive is given equal weight in cumulative counts.
- Target sectors:
  - Majority of trade-distorting measures target products from predefined lists of strategic sectors.
  - Large share target military/civilian dual use products and advanced technologies and upstream inputs such as critical minerals, chemicals, steel, and aluminum.
  - Low carbon technology products do not account for a prominent share despite climate mitigation motives; many climate-motivated measures target energy intensive industries (e.g., steel and aluminum).
  - Policies targeting critical minerals and industrial raw materials have become more salient in recent years.
- Instruments by income group and changes over time:
  - Main instrument overall: subsidization.
  - Financial grants are most common among EMDEs; state loans and trade finance dominate subsidies in AEs.
  - Public procurement localization measures are more common among AEs (accounting for over 5 percent of all trade distorting industrial policies).
  - Import restricting measures play an important role in EMDEs (18 percent of all trade distorting industrial policies).
  - Convergence between AEs and EMDEs (2009-2023):
    - Share of subsidies in AEs decreased from 84 to 75 percent.
    - Share of subsidies in EMDEs rose from 56 to 71 percent.
    - Share of trade measures in AEs increased from 3 to 8 percent.
    - Share of trade measures in EMDEs decreased from 27 to 18 percent.
    - Localization policies increased from 7 to 8 percent in AEs and from 2 to 6 percent in EMDEs.
- Instrument mixes by motive (comparisons before 2020 and after 2019):
  - Climate-motivated measures:
    - Domestic subsidies used in 89 percent of cases before 2020 and 92 percent thereafter.
  - National security/geopolitical-motivated measures:
    - Export barriers increased from 7 percent before 2020 to 22 percent after.
    - Localization or public procurement dropped from 30 to 9 percent.
    - Import barriers dropped from 28 to 11 percent.
    - FDI measures dropped from 24 to 11 percent.
    - Measures in the ‘other’ category rose from 5 to 37 percent.
  - Supply chain resilience-motivated measures:
    - Domestic subsidies used in 82 percent before 2020 and 76 percent after.
    - Export barriers increased from 3 percent to 13 percent.
  - Strategic competitiveness-motivated measures:
    - Domestic subsidies increased from 63 percent to 75 percent.
    - Export subsidies decreased from 27 percent to 11 percent.
- Opportunity cost and heterogeneity across economies:
  - Analysis using World Development Indicators for 77 economies’ tax bases and GNI relative to regional peers (last year available) shows:
    - Nearly two-thirds of economies had relatively small tax bases and market sizes, raising the opportunity cost of corporate subsidies and blunting likely effectiveness of localization measures.
    - 16 percent of economies in this sample (the green area) had larger markets regionally and states with potentially deep pockets.
    - 20 percent of economies in this sample (the yellow area) may be tempted to compete with big regional rivals but do so at a structural disadvantage.
- Extensive margin: jurisdictions using trade-distorting policies:
  - The share of jurisdictions deploying trade distorting industrial policies increased from around 56 percent in 2009 to 63 percent in 2023, reaching up to 75 percent in 2020.
  - Quarterly shares of jurisdictions are reported as three-quarter moving averages and can take values 0-1 for each instrument separately.

*IMF Working Papers — Industrial Policy since the Great Financial Crisis (content from H-NIPO / NIPO descriptions and Structural Trends: 2009-2023).*

### 2022. This change is most notable for subsidies: in 2009 only 36 percent of countries employed trade distorting

### IMF WORKING PAPERS Industrial Policy since the Great Financial Crisis

### Aggregate trends in industrial policy instruments and motives
- Subsidies:
  - In 2009 only 36 percent of countries employed trade distorting subsidies while in 2023 this had risen to 59 percent.
  - Among AEs, the share employing subsidies rose from 81 to 90 percent, reaching 100 percent in 2020.
  - Among EMDEs, the share employing subsidies rose from 36 to 57 percent, reaching 66 percent in 2020.
- Import and export measures:
  - Have remained relatively stable between 26 and 32 percent (aggregate).
  - Among AEs, import and export restrictions were used by 19 percent in 2009 and increased to 35 percent in 2023, peaking at 74 percent in 2020.
  - Among EMDEs, the share employing import and export restrictions has stayed level at 57 percent with a peak of 77 percent in 2020.
- Localization policies:
  - Doubled from 6 percent to 12 percent (aggregate).
  - Among AEs increased from 6 percent in 2009 to 16 percent by 2023.
  - Among EMDEs increased from 18 percent to 23 percent.
- Motives:
  - Competitiveness and climate remained the most frequent motives throughout the sample.
  - New-style motives—supply chain resilience, national security, and geopolitical concerns—begin to emerge around 2020.

### Empirical approach and key explanatory variables
- Dependent variable:
  - Change in implemented trade distortive industrial policy measures at the country c × product p (six-digit HS subheading) level.
- Benchmark specification (Equation (1)):
  - ∆IPMeasures_{c,t,p} = α_{ct} + γ_{pt} + θ_{cp} + β X_{p,c,t-1} + ε_{c,t,p}
- Key explanatory variables in X_{p,c,t-1}:
  - Import-weighted geopolitical distance (GPD), constructed using IPD from Bailey, Strezhnev, and Voeten (2017) and import shares.
  - Export-weighted industrial policy measures by other countries (XWIP), weighted by export shares.
  - Herfindahl–Hirschman index (HHI) of a country’s import concentration for product p.
  - Change and level of Balassa revealed comparative advantage (RCA).
  - Count/stock of past industrial policies (Count of IPs).

### Main empirical findings (country-product level; rolling six-year subperiods)
- Structural break around 2020:
  - The aggregate structural break around 2020 is also present at the country-product level; several key coefficients exhibit sign reversals around 2020.
- Geopolitical exposure (GPD):
  - Had a significant negative coefficient until the 2012-2018 period, became insignificant, and turned positive and significant in recent years.
  - Interpretation: countries previously avoided targeting products imported from geopolitical rivals; in recent years they increased support to those products (consistent with derisking after COVID-19 and Russia’s War in Ukraine).
  - Example coefficient path (Table 2, rolling windows): -0.0588***, …, 0.0433***, 0.0298** (values shown in the table).
- Export exposure to other countries’ industrial policies (XWIP):
  - Negative correlation until about 2014-2020 / 2015-2021, then flipped to positive and significant afterwards.
  - Interpretation: shift from avoiding products targeted by others to a tit-for-tat dynamic where countries introduce measures in products where large exporters have a large stock of measures.
  - Example coefficient path (Table 2): -0.3109***, …, 0.4197***, 0.4523***.
- Import concentration (HHI):
  - Consistently positive and significant in all periods.
  - Interpretation: policies target products with concentrated import sources, consistent with supply-chain derisking or supporting oligopolistic “high rent” industries.
  - Example coefficient path (Table 2): 0.0688***, …, 0.0950***, 0.0879***.
- Revealed comparative advantage (RCA):
  - Change in RCA generally not statistically significant in aggregate, but the level of RCA becomes significant in recent periods and exhibits a sign reversal between 2014-2020 and 2016-2022.
  - Interpretation: initially countries supported products with a prominent export presence (established industries with lobbying power); recently they have switched to targeting relatively uncompetitive or new products.
- Stock of past industrial policies (Count of IPs):
  - Significant sign reversal from negative to positive between the periods 2012-2018 and 2014-2020.
  - Interpretation: previously countries slowed new interventions as their stock grew; in recent years countries concentrate more support on sectors with a large existing industrial policy footprint.

### Heterogeneity by income group (AE vs EMDE)
- Aggregate (2009-2023) results can mask changing dynamics; pre- and post-2020 splits highlight differences.
- Geopolitical exposure × income group:
  - For EMDEs: negative coefficient switched from significant to insignificant post-2020.
  - For AEs: positive and significant in both pre- and post-2020 in several specifications.
  - Implication: the aggregate shift toward positive geopolitical exposure is driven predominantly by EMDE dynamics.
- Export exposure × income group:
  - No large differences by income class for the sign reversal; the coefficient turns from negative to insignificant post-2020 across groups.
- HHI × income group:
  - Import concentration positive generally; in the second period concentration is no longer significant for EMDEs, indicating less attention to derisking or oligopolistic sectors for EMDEs in that period.
- RCA dynamics by income group:
  - Change in RCA: positive and significant for AEs; for EMDEs positive and insignificant in the first period and negative and significant in the second period.
  - Level of RCA: consistently insignificant for each income class in aggregate regressions.
- Count of IPs × income group:
  - For EMDEs the coefficient is consistently negative and significant in the second period.
  - For AEs the coefficient turns from negative and significant to positive and significant.
  - Interpretation: acceleration and increased concentration of support to previously targeted sectors is driven predominantly by AEs.

### Differences across stated motives (Table 4 results)
- Motive groups: climate, competitiveness, and combined national security / GVC resilience (NS/GVC). NS/GVC also analyzed for intermediate products.
- Geopolitical exposure by motive:
  - Becomes positive and significant post-2020 for climate and NS/GVC motives but not for competitiveness.
  - For NS/GVC this holds for all products and the subsample of intermediate products.
- Export exposure by motive:
  - Negative and significant in both periods for climate and competitiveness motives.
  - Positive and significant for NS/GVC measures—this drives the aggregate sign reversal identified earlier.
- RCA and motive:
  - Changes in RCA post-2020: positive and significant for NS/GVC; negative and significant for competitiveness.
  - Interpretation: NS/GVC measures more often applied in sectors of growing competitiveness, while competitiveness-motivated measures are applied where comparative advantage is sliding.
  - Level of RCA post-2020: positive and significant for climate and competitiveness motivated measures.
- Stock of past measures by motive:
  - Negative and significant association for climate and competitiveness motives in both periods.
  - For NS/GVC, association turns from negative to positive post-2020, implying increasing concentration of NS/GVC measures in sectors already targeted.

### Conclusions and research implications
- 2020 marks a structural shift toward increased use of industrial policy in both AEs and EMDEs.
- Subsidies are the main policy instrument in both AEs and EMDEs; share of countries implementing them has risen markedly since the GFC.
- Trade measures are more commonly deployed by EMDEs.
- Empirical evidence:
  - Geopolitical exposure, export exposure to other countries’ policies, and the stock of domestic policies all changed sign around 2020—consistent with derisking, tit-for-tat dynamics, and increased sectoral concentration of policies.
  - Revealed comparative advantage switched from a positive to a negative correlation (driven by AEs), indicating a shift toward less established sectors.
  - NS/GVC-motivated measures account for many observed shifts in behavior.
- Dataset contribution and future research directions:
  - The historical extension of the New Industrial Policy Observatory (NIPO) provides a long-running inventory useful for studying dynamic impacts of non-market interventions on comparative advantage, the effectiveness of measures in meeting non-trade objectives, welfare implications under specific conditions, and international spillovers relevant to WTO reform debates.
  - Suggested future work: exploit the over 15-year horizon to assess dynamic impacts, motive-specific success, net welfare effects, and cross-country spillovers.

*IMF WORKING PAPERS Industrial Policy since the Great Financial Crisis*

### Annex I. Methodology for Global Trade Alert Data

### Annex I. Methodology for Global Trade Alert Data

### Collection
- The GTA initiative documents credible announcements of meaningful and unilateral changes by governments that affect the relative treatment of foreign versus domestic commercial interests.
- The dataset begins in November 2008.
- Emphasis: unilateral policy changes (to distinguish from regional, plurilateral, and multilateral trade agreements).
- Foreign commercial interests considered: trade in goods and services, investment, and labor force migration.
- Over 60 different types of commercial policy intervention—including subsidies—are documented.
- The GTA initiative does not track changes in Technical Barriers to Trade (TBT) and Sanitary and Phytosanitary Measures (SPS).
- Each GTA database entry includes:
  - Implementing jurisdiction.
  - Direction of the change (distortive or liberalizing).
  - Announced policy instrument.
  - Announcement date and, where available, implementation date.
  - Sectors and products covered.
  - For measures affecting cross-border trade in goods: potentially affected trading partners (identified based on official United Nations trade flow data).
- Source basis and review:
  - Each entry is based on the official statement by the responsible institution wherever possible.
  - All entries undergo a two-stage review process before publication.
  - Each announcement documented includes at least one new and credible public declaration for change in market conditions at home or abroad.
  - An announcement may involve several unilateral changes in policy interventions.
  - As of this writing, over 60,000 policy interventions have been documented since the GTA initiative began.

### Entry criteria — seven conditions for a new GTA database entry
- 1) Unilateral Action: the intervention shall be a deliberate action that tilts the playing field to benefit or harm foreign commercial interests. Interventions that are bi-, pluri- or multilaterally agreed are beyond the scope.
- 2) Relative Treatment Test: the intervention must alter the relative treatment of domestic commercial interests vis-à-vis foreign competitors. A measure is deemed distortive if it discriminates against foreign commercial entities in favor of at least one rival with operations in the implementing jurisdiction.
- 3) Meaningful Change: the intervention is likely to meaningfully change international commercial flows.
- 4) Credible Action: the intervention must be implemented already or its future implementation date is enacted and known.
- 5) Absence of uncontested higher motive: multilateral measures with a codified set of goals that are superior to the preservation of seamless international commerce are not included. Such codification can happen through international treaties, agreements or resolutions in the public domain.
- 6) One announcement, one entry: interventions with the same announcement are to be reported in the same GTA database entry.
- 7) GTA monitoring period: the meaningful change has to be announced on or after 1 November 2008.

---

### Annex III. Classifying Industrial Policy Motives using a Large Language Model Ensemble Approach

### Rationale and training data
- Policy descriptions from GTA and NIPO contain intentions but attributing motives is non-trivial.
- Traditional bag-of-words methods can miss contextual nuance; LLMs allow more contextual interpretation.
- Only a subset of GTA measures (those in NIPO) contain human annotations identifying stated motives; this annotated subset forms the training and validation dataset.
- RoBERTa is chosen as the base model.

### Pretrain-finetune paradigm and ensemble approach
- Adopt pretrain-finetune paradigm to leverage LLM strengths for classification.
- To address instability due to randomization during training, train RoBERTa ten times with randomized classification head initialization and randomized batch order, then average probabilities across the ensemble.

### Algorithm in Detail
- Step 0 – Model Construction:
  - RoBERTa generates 768-dimensional vector representations for each input token.
  - Use the 768-dimensional hidden state as input to a custom classification head.
  - Classification head architecture:
    - Fully connected layer (768 × 768).
    - GeLU activation function.
    - Dropout layer (dropout rate: 0.1).
    - Final classification layer (768 × 2).
    - Softmax applied at output to generate probability distributions over target classes.
  - Full finetuning: all model weights—including RoBERTa’s pre-trained weights and classification head weights—are updated during training.
- Step 1 – Training/Testing set split:
  - Randomly split labeled NIPO dataset into training and testing sets.
  - Use a class-stratified 80% sample for training and reserve the remaining 20% for testing.
  - The train/test split is performed each time of fine-tuning with a different seed number.
  - The test dataset is withheld until the entire algorithm is complete for final evaluation.
- Step 2 – Finetuning:
  - Supervised learning with cross-entropy loss:
    - L = − ∑_{i=1}^{N} w_i ∙ y_i log(ŷ_i)  (notation preserved from source).
    - y_i is the true label, ŷ_i is the predicted probability for class i, w_i is the class weight, N is the number of classes.
  - Incorporate class weights w_i to account for label imbalance.
  - Training regimen and hyperparameters:
    - Follow Mosbach et al. (2021) best practices: small learning rate with bias correction and many iterations.
    - Optimizer: AdamW.
    - Learning rate: 1×10^−5.
    - Train for 10 epochs.
    - Weight decay: 0.01.
    - Linear learning rate scheduler with warm-up; warm-up proportion set to 10% of total training steps.
  - Repeat this step ten times to produce ten finetuned models for each stated motive.
- Step 3 – Ensemble and Production:
  - Apply algorithm to all IPs from GTA dataset.
  - For each policy, feed policy title and description to each of the ten finetuned models to obtain ten sets of probabilities per stated motive.
  - Compute the weighted average of the probabilities.
  - Classify the policy as having the stated motive if the weighted probability exceeded 60%.

### Performance
- Average precision and recall across the 10 fine-tuned models (assessed with the test datasets):
  - Without motive:
    - precision 0.95
    - recall 0.93
  - Strategic competitiveness (with motives):
    - precision 0.80
    - recall 0.83
  - GVC Resilience:
    - precision 0.83
    - recall 0.75
  - Geopolitical concerns or national security:
    - precision 0.90
    - recall 0.93
  - Climate change mitigation:
    - precision 0.84
    - recall 0.87
- Summary statement from source: the algorithm performs best for policies without motive, achieves a precision above 80% for all motives and recall above 80% for three out of four motives.

---

### Annex IV. Additional Descriptive Statistics

- Table A1. Frequency of Policy Instrument by Stated Motive (Share (%) 2009-2019; Share (%) 2020-2023)
  - Climate Change Mitigation:
    - Domestic subsidy: 89; 92
    - Export barrier: 0; 0
    - Export subsidy: 6; 3
    - FDI: 0; 0
    - Import barrier: 0; 0
    - Localization or Procurement: 5; 5
    - Other: 0; 0
  - National Security and/or Geopolitical Concern:
    - Domestic subsidy: 6; 10
    - Export barrier: 7; 22
    - Export subsidy: 0; 0
    - FDI: 24; 11
    - Import barrier: 28; 11
    - Localization or Procurement: 30; 9
    - Other: 5; 37
  - Resilience/Security of Supply Chains (Non-food):
    - Domestic subsidy: 82; 76
    - Export barrier: 3; 13
    - Export subsidy: 5; 3
    - FDI: 0; 1
    - Import barrier: 4; 2
    - Localization or Procurement: 5; 4
    - Other: 1; 0
  - Strategic Competitiveness:
    - Domestic subsidy: 63; 75
    - Export barrier: 0; 1
    - Export subsidy: 27; 11
    - FDI: 1; 0
    - Import barrier: 4; 7
    - Localization or Procurement: 5; 6
    - Other: 0; 0

*International Monetary Fund — Annex I. Methodology for Global Trade Alert Data (from source PDF).*

---


_Source: https://www.imf.org/-/media/files/publications/wp/2025/english/wpiea2025222-source-pdf.pdf_