## fd-dec25-lowres - 2024. Each dot represents a country’s

## Source details

**Canonical URL:** [fd-dec25-lowres - 2024. Each dot represents a country’s](https://www.imf.org/-/media/files/publications/fandd/article/2025/12/fd-dec25-lowres.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/fandd/article/2025/12/fd-dec25-lowres.pdf.md)
- [Structured JSON version](/-/media/files/publications/fandd/article/2025/12/fd-dec25-lowres.pdf.json)

---

### The explosion of data and the need for new measurement mindsets
- Expansion from a “small data” world to a “big data” universe requires rethinking what is measured and how.
- Historical analogy: GE’s 1990 MRI software update showed professionals may resist more accurate data because skills and institutions were tuned to older, compressed representations.
- Benefits of new data:
  - Greater accuracy, speed (quasi real time), and granularity (down to small segments or individuals).
  - Ability to change what gets measured, not just measurement methods.
- Risks and caveats:
  - Much new data is private-sector generated; harnessing it requires validation, collaboration, and attention to biases and continuity.

### Alternative indicators (alt-data): opportunities and limitations
- Private data examples and scale:
  - ADP handles one in six US workers.
  - PriceStats tracks changes in 800,000 daily prices from among 40 million products in 25 economies.
  - LinkedIn’s “economic graph” measures:
    - 1.2 billion people
    - 67 million companies
    - 15 million jobs
    - 41,000 skills
    - 133,000 schools
  - Carlyle manages 277 companies with 730,000 employees.
- Use cases where alt-data proved valuable:
  - Early warning during the 2008 downturn and in the COVID-19 pandemic (mobility, visits to retailers, supply-chain stress, shipping container movements).
  - Argentina’s official inflation credibility issues led The Economist to use PriceStats in the early 2010s.
- Limitations:
  - Selection biases (e.g., LinkedIn skews professional; ADP omits informal sector).
  - Private indices may be discontinued or methodologically changed (Intuit’s QuickBooks small-business index discontinued in 2015 and relaunched in 2023 with a different methodology).
  - Continuity and access risks if providers stop sharing data or raise prices.

### Privacy, encryption, and ethical considerations
- Techniques to analyze sensitive data without exposing raw records: federated learning, homomorphic encryption, secure multiparty computation, and differential privacy.
- Potential future metrics could approach “ground truth” (e.g., measures of individual misery using spending switches, missed payments, health and emotional biomarkers) but carry profound and controversial privacy implications.

### Central banks and nontraditional data in monetary policy (Claudia Sahm)
- Nontraditional data are complements, not substitutes, to traditional representative surveys.
- Timeliness advantage: private data releases can be days or weeks versus government statistics published weeks or months later.
  - Example: Fed’s internal weekly ADP employment estimates signaled declines in late March 2020 more than a month before the BLS monthly report.
- Granularity for policy analysis:
  - Daily price indices built from online data for 350,000 goods (including country of origin and tariff rates) show imported consumer goods’ prices rose faster than domestically produced goods relative to pre-tariff trends; overall effects were relatively modest.
- Improving initial estimates and model inputs:
  - Weekly tax filings for employer identification numbers provide reliable forecasts of business formation and could align birth-death models to improve payroll employment estimates.
  - A Federal Reserve Bank of Chicago initiative blends official and alternative data (Indeed, Lightcast, Google searches) to estimate current-month unemployment; project is in early stages.
- Distributional and impact analysis:
  - Household-level credit, bank accounts, and administrative records revealed distributional effects of monetary policy (e.g., discrepancies in refinancing during COVID-19 across racial and income groups).
- Fed-specific challenges:
  - Limited public availability of private data reduces transparency and external verifiability.
  - Data-provider-specific artifacts (e.g., client growth) require benchmarking against representative sources like the Economic Census.

### Modernizing macroeconomic measurement (Rebecca Riley)
- Digital economy and intangibles:
  - Data-driven technologies and intangibles (software, marketing databases, organization capital) are increasingly central but undercounted in GDP and productivity measures.
  - Conference Board research: 50% of intangible investments in advanced economies are essentially data investments.
- Specific measurement issues:
  - Profit shifting by multinationals (intellectual property ownership, revenue location) distorts national output and trade statistics.
  - Quality adjustment problems: failure to adjust for rapid quality improvements can misstate productivity (example: UK telecommunications quality-adjusted price adjustments reduced the estimated productivity slowdown by 0.25 percentage point).
  - Free digital services are widely consumed but often not captured in household consumption; experimental UK work estimated the nominal value of digital leisure services at 8 percent of nominal GDP.
  - Household production of digital goods and services requires time-use data to assess scale.
- Harnessing private data for official statistics:
  - Point-of-sale and online retailer data have been progressively incorporated by national agencies (The Netherlands, Australia, Canada, and emerging work in the UK).
  - Technical advances in handling large-scale messy data enable new methods (research by Kevin Fox and colleagues).
- Institutional and resource challenges:
  - Upfront costs, capability building, legal and technical frameworks, data-sharing agreements, trusted institutions, and international coordination are required.
  - The 2025 revision to the System of National Accounts (SNA) and updates to the IMF’s balance of payments manual are starting points that need widespread adoption by statistical agencies.
- Recommended collaborative approaches:
  - Strengthen partnerships among national statistical agencies, private data owners, government officials, and academics.
  - Invest in data-processing technologies, proofs of concept, new statistical methods, and data exploration.
  - Maintain statistical rigor, impartiality, transparency, and equal access to trusted official statistics.

### Key numeric facts and illustrative figures preserved from the source
- ADP handles one in six US workers.
- PriceStats tracks changes in 800,000 daily prices from among 40 million products in 25 economies.
- LinkedIn’s “economic graph” measures:
  - 1.2 billion people
  - 67 million companies
  - 15 million jobs
  - 41,000 skills
  - 133,000 schools
- Carlyle manages 277 companies with 730,000 employees.
- Intuit discontinued its QuickBooks small-business index in 2015 and relaunched it in 2023.
- Conference Board estimate: 50% of intangible investments in advanced economies are data investments.
- Experimental UK estimate: nominal value of digital leisure services at 8 percent of nominal GDP.
- Research finding: quality-adjustment methodological changes shaved 0.25 percentage point off estimated UK productivity slowdown for the decade after the 2008 financial crisis.
- Historical references preserved as in source: 1974, 1980, 1990, 2008, 2011, 2014, 2015, 2020, 2023, 2025.

### The hidden price of data — valuation approaches and implications
- Core insight:
  - Data functions as an ambient, invisible asset produced as a by-product of transactions; observable monetary prices for digital goods bundle payment for the good and an implicit discount for transaction data.
- Five measurement approaches to value data:
  - Market prices approach: uses data marketplaces (examples: Snowflake, Datarade); limitation: traded data is not representative of the most valuable data held by firms.
  - Revenue approach: treats data like a productive asset; requires counterfactual modeling of profits without specific data.
  - Complementary inputs approach: infers value from resources devoted to managing and exploiting data (labor, computing).
  - Correlated behavior approach: measures how data improves decisions by estimating covariance between actions and payoffs.
  - Cost-accounting approach: adds up bills and counts purchased data sets as assets per UN SNA; difficulty: most data is bartered and implicit discounts rarely appear on accounting records.
- Policy and measurement implications:
  - Unbundling the price for the right to use transaction data and the price for a private transaction would reveal implicit data discounts.
  - A complete accounting of data would need to impute the value of implicit discounts given to consumers.
  - No single measurement approach is sufficient; a toolkit of methods is necessary.

### Inside the AI-led resource race — findings, projections, and policy challenges
- Physical inputs and infrastructure:
  - Hyperscale data centers have power needs in the tens of megawatts; planned projects include a 5-gigawatt campus in Abu Dhabi spanning 10 square miles.
- Electricity and energy demand:
  - Data centers already use about 1.5 percent of global electricity supply (comparison: roughly the same as the United Kingdom).
  - The IEA expects data center demand to more than double by 2030, with AI responsible for much of the increase.
  - AI accounts for less than a tenth of added power demand this decade globally, but national impacts vary:
    - In the US and Japan, data centers could account for nearly half of new demand by 2030.
    - In Ireland, data centers already use more than a fifth of the country’s electricity.
  - Local concentrations produce acute grid stress (example: northern Virginia data centers consume about one-quarter of the state’s power).
- Corporate responses and energy sourcing:
  - Microsoft, Amazon, and Google have signed multibillion-dollar power purchase agreements and are major corporate buyers of renewable energy.
  - Firms experiment with on-site generation and alternative technologies: Microsoft explored nuclear (small modular reactors), Google backs advanced geothermal, Amazon tests hydrogen for backup power.
  - Despite renewable investments, US data center demand still leans heavily on natural gas and could add to emissions.
- Efficiency, Jevons paradox, and chips:
  - New chips (example: Nvidia’s Blackwell processors) and TPUs improve operations per watt.
  - Software innovations (example: China’s DeepSeek, released in January 2025) were trained at a fraction of cost and energy of comparable models.
  - Efficiency gains may spur greater use (Jevons paradox), potentially increasing overall resource consumption.
- Semiconductors and geopolitical concentration:
  - Training state-of-the-art models requires thousands of specialized chips, most designed by Nvidia and manufactured predominantly by TSMC in Taiwan Province of China.
  - The US has restricted advanced chip exports to China and subsidized domestic fabrication; China and other jurisdictions are scaling domestic capabilities.
- Mineral and material demands (projections to 2030):
  - Data centers could consume:
    - more than half a million metric tons of copper each year,
    - 75,000 tons of silicon each year.
  - These levels could lift data centers’ share of global demand to 2 percent (IEA estimate).
  - For gallium, data centers could account for more than a tenth of total demand.
  - China controls 80–90 percent of global refining of silicon, gallium, and rare earths and imposed export restrictions in 2023 and new curbs since late 2024 on tungsten, tellurium, bismuth, indium, and molybdenum.
- Land, water, and siting:
  - Two-thirds of new US centers since 2022 have been built in water-stressed regions.
  - Policy examples: Ireland froze new projects in 2022 unless projects could generate their own power; Singapore halted approvals in 2019 and now allows facilities only under strict efficiency rules.
- Policy challenges (enumerated):
  - Forecast uncertainty: published forecasts for data center demand for 2030 diverge widely; the highest published estimate is nearly seven times the lowest.
  - Rapid build-out: governments must expand electricity systems quickly without overbuilding or locking in fossil fuels.
  - Transparency gap: “There is little public reporting from the industry on data center use of electricity, water, or minerals.”
  - Sustainability and equity: risk of repeating boom-and-bust commodity cycles and skewing benefits toward the rich world.
- Policy levers and potential outcomes:
  - If managed well, the AI boom could accelerate clean energy investment and foster more resilient supply chains.
  - If mismanaged, it risks locking in new emissions and deepening resource dependence.
  - Governments must integrate power plants, grids, water, and minerals into digital policy planning.

### India: modernizing official statistics (interview with Saurabh Garg)
- Scope and staffing:
  - Population reference: "1.4 billion."
  - Staffing: "some 5,000 full-time staff members at the central level and over 6,000 field investigators and supervisors during major survey operations across the nation’s 28 states and 8 territories."
- Technology and data collection:
  - Door-to-door collection remains, but data collection is tablet based.
  - AI chatbots are being added to tablets to answer enumerators’ questions immediately.
  - All tablet data is uploaded directly to a background portal, called "e‑SIGMA."
  - Website improvements: mobile app, more infographics, direct unit-level data downloads, and training videos.
- Incorporating alternative data and quality assurance:
  - Exploring incorporation of "e-commerce, scanner, mobile, satellite, and other alternative data" while emphasizing rigorous scientific methods and comprehensive quality control.
- Labor statistics enhancements:
  - Since January 2025: produced a "monthly report for both the urban and rural sectors."
  - Also have a "quarterly report" with more sectoral and employment detail.
  - Sample size: "We have nearly doubled the sample size."
- Harmonization and metadata:
  - Registry of government data sets, a "national metadata structure," alignment with international systems, unique identifiers for organizations and locations, and methods to reconcile discrepancies.
- Adoption of international standards and rebasing:
  - Adopted the United Nations Fundamental Principles of Official Statistics in 2016.
  - "We’re in the process of rebasing our national accounts" and incorporating SNA 2025 changes into rebased methodology and new guidelines to be released over the next couple of years.
- Challenges:
  - Need for more granular data at district, subdistrict, and village levels implies increasing sample sizes and margin of error.
  - Frequency demands: users "aren’t willing to wait a year for data. Even a month is too long."
  - Ensuring cooperation and maintaining privacy and credibility in a social media age.

### Nowcasting and new data sources in developing economies
- Concept and value:
  - Nowcasting uses real-time and high-frequency data to predict the present and near future of economic activity; originated in meteorology and adapted to macroeconomics.
  - Provides timelier signals than official indicators during fast-moving events (e.g., the pandemic).
- Data gaps and frequency challenges:
  - "Advanced economies and most emerging markets publish GDP quarterly. But about a third of countries in the world have only annual GDP."
  - First quarterly GDP estimates in advanced economies typically appear about a month after quarter end—"two months in some major emerging market economies"—and in developing economies it can take "more than three months."
- Case examples:
  - Rwanda: weekly economic activity index during COVID using central bank and tax-authority electronic billing data.
  - Kenya: nowcasting allows central bank to start gauging the quarter "after only a week" using private consumer spending data; remittance data after two weeks; trade, money supply, tourism, and electricity data in about 40 days.
  - South Africa: commercial agricultural data lead CPI by one to three months across components.
- Best practices and cautions:
  - Nowcasting complements but does not replace rigorous official data collection.
  - "Start with the data": identify high-quality or proxy data before selecting techniques; simple models often perform best.
  - New data sources can be noisy or nonrepresentative; models can fail when structural conditions change.
  - Broader concerns: privacy, transparency, legitimacy, governance, and capacity constraints.

### Case studies and technological innovations (selected)
- Democratic Republic of the Congo (DRC): IMF-assisted central bank nowcasting combining traditional indicators (copper and cobalt production and prices, money supply) with nontraditional inputs (satellite night-light intensity, Google search trends) for a quarterly projection model; mining sector contributes "about a third of GDP."
- Guinea-Bissau (blockchain for wage bill): May 2024 deployment (with IMF, EY, donors) to manage public wage bill; "The wage bill fell to half of tax revenue in 2024." Platform to expand to "all 26,600 public officials and 8,100 pensioners nationwide."
- Kenya: nowcasting models approximate growth ahead of official releases; Central Bank of Kenya and IMF cooperation on nowcasting and bimonthly surveys.
- Madagascar (AI customs): October introduction of "agentic AI" to detect inconsistencies signaling fraud; automated frontline tasks freed experts for complex cases.
- Broader finding: IMF technical assistance supports capacity development and adoption of these tools across countries.

### Policy recommendations, operational steps, and governance
- For national statistical systems:
  - Make administrative and ministry data machine readable.
  - Maintain a registry of government datasets and a national metadata structure; use unique identifiers for interoperability.
  - Reconcile discrepancies among administrative datasets via formal methods.
  - Uphold the United Nations Fundamental Principles of Official Statistics.
  - Hold stakeholder consultations and publish detailed documentation for large-sample surveys.
- For data innovation and nowcasting:
  - "Start with the data": identify high-quality domestic or proxy datasets before applying complex models.
  - Use alternative high-frequency data (billing, retail, satellite, search trends) while recognizing limitations.
  - Invest in analytical capacity: economists, statisticians, data scientists, and computing resources.
  - Balance frequency and granularity with statistical reliability; acknowledge trade-offs between sample size, granularity, and margin of error.
- For governance and public trust:
  - Ensure transparency, documentation, and independent scrutiny of methodologies.
  - Address privacy concerns and maintain public cooperation through clear safeguards.
  - Protect statistical agencies from political influence by adhering to established principles and ensuring professional independence.

*Source: fd-dec25-lowres - 2024. Each dot represents a country’s (content as provided).*

### 2024. Each dot represents a country’s

### fd-dec25-lowres - 2024. Each dot represents a country’s

### The explosion of data and the need for new measurement mindsets
- The variety, frequency, and granularity of data sources have expanded from a “small data” world to a “big data” universe, requiring economists to rethink what is measured and how.
- Historical analogy: GE’s 1990 MRI software update showed professionals may resist more accurate data because skills and institutions were tuned to older, compressed representations.
- Benefits of new data:
  - Greater accuracy, speed (quasi real time), and granularity (down to small segments or individuals).
  - Ability to change what gets measured, not just measurement methods.
- Risks and caveats:
  - Private sector generates much of the new data; harnessing it requires validation, collaboration, and attention to biases and continuity.

### Alternative indicators (alt-data): opportunities and limitations
- Examples of private data and capabilities:
  - Payroll processor ADP handles one in six US workers and supplies monthly jobs data used to supplement US Bureau of Labor Statistics reports.
  - PriceStats tracks changes in 800,000 daily prices from among 40 million products in 25 economies.
  - LinkedIn’s “economic graph” measures the work activities of 1.2 billion people, 67 million companies, 15 million jobs, 41,000 skills, and 133,000 schools.
  - Carlyle manages 277 companies with 730,000 employees and provided employment trends when official data was unavailable.
- Use cases where alt-data proved valuable:
  - Early warning during the 2008 downturn and in the COVID-19 pandemic (mobility, visits to retailers, supply-chain stress, shipping container movements).
  - Argentina’s official inflation credibility issues led The Economist to use PriceStats in the early 2010s.
- Limitations of alt-data:
  - Often data exhaust by-product of business processes and carries selection biases (e.g., LinkedIn skews professional; ADP omits informal sector).
  - Private providers may discontinue indices (example: Intuit’s QuickBooks small-business index discontinued in 2015 and relaunched in 2023 with a different methodology).
  - Continuity and access risks: a private company could stop sharing data or raise prices.

### Privacy, encryption, and ethical considerations
- Advanced techniques exist to analyze sensitive data without exposing raw records: federated learning, homomorphic encryption, secure multiparty computation, and differential privacy.
- Potential future metrics could approach “ground truth” (e.g., measures of individual misery using spending switches, missed payments, health and emotional biomarkers), but privacy implications are profound and controversial.

### Central banks and nontraditional data in monetary policy (Claudia Sahm)
- Role of nontraditional data for the Fed:
  - Nontraditional data are complements, not substitutes, to traditional representative surveys.
  - Timeliness advantage: private data releases can be days or weeks vs. government statistics published weeks or months later.
  - Example: Fed’s internal weekly ADP employment estimates signaled declines in late March 2020 more than a month before the BLS monthly report.
- Granularity for policy analysis:
  - Daily price indices built from online data for 350,000 goods including country of origin and tariff rates (work by Alberto Cavallo and collaborators) show imported consumer goods’ prices rose faster than domestically produced goods relative to pre-tariff trends; overall effects were relatively modest.
- Improving initial estimates and model inputs:
  - Weekly tax filings for employer identification numbers provide reliable forecasts of business formation and could align birth-death models to improve payroll employment estimates.
  - A Federal Reserve Bank of Chicago initiative blends official and alternative data (Indeed, Lightcast, Google searches) to estimate current-month unemployment; project is in early stages.
- Distributional and impact analysis:
  - Household-level credit, bank accounts, and administrative records helped show distributional effects of monetary policy (e.g., discrepancies in refinancing during COVID-19 across racial and income groups).
- Challenges specific to the Fed:
  - Limited public availability of private data reduces transparency and external verifiability.
  - Data-provider-specific artifacts (e.g., client growth) require benchmarking against representative sources like the Economic Census.

### Modernizing macroeconomic measurement (Rebecca Riley)
- Digital economy mismatch with existing metrics:
  - Data-driven technologies and intangibles (software, marketing databases, organization capital) are increasingly central but undercounted in GDP and productivity measures.
  - Conference Board research: fully half of intangible investments in advanced economies are essentially data investments that economic accounts are only starting to include.
- Specific measurement issues:
  - Profit shifting by multinationals (intellectual property ownership, revenue location) distorts national output and trade statistics.
  - Quality adjustment problems: digital products and services often improve quality rapidly; failure to adjust can misstate productivity (example: UK telecommunications quality-adjusted price adjustments reduced the estimated productivity slowdown by 0.25 percentage point).
  - Free digital services (search engines, social media, open-source software) are widely consumed but often not captured in household consumption; experimental work estimated the nominal value of digital leisure services in the UK at 8 percent of nominal GDP.
  - Household production of digital goods and services requires time-use data to assess scale.
- Harnessing private data for official statistics:
  - Point-of-sale and online retailer data have been progressively incorporated by national agencies in The Netherlands, Australia, Canada, and emerging work in the UK to improve consumer price indices.
  - Technical advances in handling large-scale messy data are enabling new methods (research by Kevin Fox and colleagues).
- Institutional and resource challenges:
  - Upfront costs, capability building, legal and technical frameworks, data-sharing agreements, trusted institutions, and international coordination are needed to operationalize private data for public statistics.
  - The 2025 revision to the System of National Accounts (SNA) and updates to the IMF’s balance of payments manual are starting points but effective implementation requires widespread adoption by statistical agencies.
- Recommended collaborative approaches:
  - Strengthen partnerships among national statistical agencies, private data owners, government officials, and academics.
  - Invest in data-processing technologies, proofs of concept, new statistical methods, and data exploration.
  - Maintain statistical rigor, impartiality, transparency, and equal access to trusted official statistics to avoid a noisy landscape dominated by private signals.

### Key numeric facts and illustrative figures preserved from the source
- ADP handles one in six US workers.
- PriceStats tracks changes in 800,000 daily prices from among 40 million products in 25 economies.
- LinkedIn’s “economic graph” measures:
  - 1.2 billion people
  - 67 million companies
  - 15 million jobs
  - 41,000 skills
  - 133,000 schools
- Carlyle manages 277 companies with 730,000 employees.
- Intuit discontinued its QuickBooks small-business index in 2015 and relaunched it in 2023.
- Conference Board estimate: 50% of intangible investments in advanced economies are data investments.
- Experimental UK estimate: nominal value of digital leisure services at 8 percent of nominal GDP.
- Research finding: quality-adjustment methodological changes shaved 0.25 percentage point off estimated UK productivity slowdown for the decade after the 2008 financial crisis.
- Historical and timeline references preserved as in source: 1974 (MRI patent and recession), 1980 (U.S. presidential election use of “misery index”), 1990 (GE MRI update), 2008 (financial crisis), 2011 (final GDP revision), 2014 (Greenspan interview), 2015 (Intuit index discontinued), 2020 (pandemic onset), 2023 (Intuit relaunch), 2025 (SNA revision and other 2025 references).

*Source: fd-dec25-lowres - 2024. Each dot represents a country’s (content as provided).*

### references

### fd-dec25-lowres - references

### References cited
- Abdirahman, M., D. Coyle, R. Heys, and W. Stewart. 2020. “A Comparison of Deflators for Telecommunications Services Output.” Economie et Statistique/Economics and Statistics 517-518-519: 103–22.
- Corrado, C., J. Haskel, M. Iommi, and C. Jona-Lasinio. 2022. “Measuring Data As an Asset: Framework, Methods and Preliminary Estimates.” OECD Economics Department Working Paper 1731, Organisation for Economic Co-operation and Development, Paris.
- Fox, K. J., P. Levell, and M. O’Connell. 2025. “Inflation Measurement with High Frequency Data.” Journal of Business & Economic Statistics. DOI: 10.1080/07350015.2025.2537392.
- Schreyer, P. 2022. “Accounting for Free Digital Services and Household Production: An Application to Facebook (Meta).” Eurostat Review on National Accounts and Macroeconomic Indicators (EURONA): 7–26.
- International Energy Agency (IEA). 2025. Energy and AI. Paris: Organisation for Economic Co-operation and Development and IEA.

### The hidden price of data — key findings and analytical approaches
- Core insight:
  - Data functions as an ambient, invisible asset produced as a by-product of transactions (searches, clicks, purchases).
  - Observable monetary prices for digital goods and services are net prices that bundle payment for the good and implicit payment (discount) for transaction data.
  - Bundling hides the price of data, preventing consumers from observing the implicit discount they receive for selling data.
- Economic logic and market behavior:
  - Profit-maximizing firms have an incentive to discount goods and services to generate more transactions and thereby acquire more data.
  - Consumers rarely have the option to purchase without selling data; thus, they cannot learn the value of their data over time.
- Five measurement approaches to value data (enumerated):
  - Market prices approach:
    - Uses prices from data marketplaces (examples: Snowflake, Datarade) where data sets are bought and sold.
    - Limitation: traded data is not representative of the most economically valuable data held by firms.
  - Revenue approach:
    - Treats data like a productive asset: worth whatever extra revenue it can generate.
    - Requires counterfactual modeling of profits without specific data.
  - Complementary inputs approach:
    - Infers value from resources devoted to managing and exploiting data (labor, computing power).
    - Implicit value arises because firms spend real money to use data.
  - Correlated behavior approach:
    - Measures how data improves decisions by estimating covariance between actions and payoffs (e.g., recommendation accuracy, inventory matching).
    - Interprets high covariance as evidence of valuable data.
  - Cost-accounting approach:
    - Adds up bills and counts purchased data sets as assets per the United Nations System of National Accounts.
    - Difficulty: most data is bartered (consumers “pay” with information), so implicit discounts rarely appear on accounting records.
- Policy and measurement implications:
  - Unbundling requirement (post both the price for the right to use transaction data and the price for a private transaction) would reveal the implicit data discount and enable consumers to decide whether to sell data.
  - A complete accounting of data would need to impute the value of implicit discounts given to consumers to encourage transactions and reveal data.
  - No single measurement approach is infallible or holistic; a toolkit of methods is necessary to move data from intuition to quantification.

### Inside the AI-led resource race — key findings, projections, and policy challenges
- Physical reality of AI:
  - AI depends on physical inputs: electricity, cooling water, chips, and minerals.
  - Hyperscale data centers have power needs in the tens of megawatts; planned projects include a 5-gigawatt campus in Abu Dhabi spanning 10 square miles.
- Electricity and energy demand:
  - Data centers already use about 1.5 percent of global electricity supply (comparison: roughly the same as the United Kingdom).
  - The IEA expects data center demand to more than double by 2030, with AI responsible for much of the increase.
  - AI accounts for less than a tenth of added power demand this decade globally, but national impacts vary:
    - In the US and Japan, data centers could account for nearly half of new demand by 2030.
    - In Ireland, data centers already use more than a fifth of the country’s electricity.
  - Local concentrations produce acute grid stress (examples: northern Virginia data centers consume about one-quarter of the state’s power).
- Corporate responses and energy sourcing:
  - Microsoft, Amazon, and Google have signed multibillion-dollar power purchase agreements and are major corporate buyers of renewable energy.
  - Firms are experimenting with on-site generation and alternative technologies: Microsoft has explored nuclear (small modular reactors and potential acquisitions), Google backs advanced geothermal, Amazon is testing hydrogen for backup power.
  - Despite renewable investments, US data center demand still leans heavily on natural gas and could add to emissions.
- Efficiency and Jevons paradox:
  - New chips (example: Nvidia’s Blackwell processors) and TPUs improve operations per watt.
  - Software innovations (example: China’s DeepSeek, released in January 2025) were trained at a fraction of cost and energy of comparable models.
  - Lower cost/greater efficiency may spur greater use (Jevons paradox), potentially increasing overall resource consumption.
- Semiconductors and geopolitical concentration:
  - Training state-of-the-art models requires thousands of specialized chips, most designed by Nvidia and manufactured predominantly by TSMC in Taiwan Province of China.
  - The US has restricted advanced chip exports to China and subsidized domestic fabrication; China and other jurisdictions are scaling domestic capabilities.
- Mineral and material demands:
  - By 2030, data centers could consume:
    - more than half a million metric tons of copper each year,
    - 75,000 tons of silicon each year.
  - These levels could lift data centers’ share of global demand to 2 percent (IEA estimate).
  - For gallium, data centers could account for more than a tenth of total demand.
  - China controls 80–90 percent of global refining of silicon, gallium, and rare earths and imposed export restrictions in 2023 and new curbs since late 2024 on tungsten, tellurium, bismuth, indium, and molybdenum.
  - Price spikes for many metals have prompted critical-mineral strategies in Washington, Brussels, Tokyo, and Seoul (recycling programs and alliances with resource-rich countries).
- Land, water, and siting considerations:
  - Hyperscale centers locate where cheap power, abundant water, and fast fiber converge.
  - Two-thirds of new US centers since 2022 have been built in water-stressed regions.
  - Examples of policy responses: Ireland froze new projects in 2022 unless projects could generate their own power; Singapore halted approvals in 2019 and now allows facilities only under strict efficiency rules.
  - Cold climates and proximity to transatlantic cables shape siting (examples: Norway, Iceland, Ireland).
- Policy challenges (enumerated):
  - Forecast uncertainty:
    - Published forecasts for data center demand for 2030 diverge widely; the highest published estimate is nearly seven times the lowest.
  - Rapid build-out:
    - The pace of building leaves little time for certainty; governments must expand electricity systems quickly without overbuilding or locking in fossil fuels.
  - Transparency gap:
    - “There is little public reporting from the industry on data center use of electricity, water, or minerals.” Greater disclosure is needed for regulators, utilities, and communities.
  - Sustainability and equity:
    - Expanding grids and supply chains without environmental and social safeguards risks repeating boom-and-bust commodity cycles.
    - Benefits of the AI boom may skew toward the rich world if developing economies remain raw-material suppliers and face higher implied energy and capital costs.
- Policy levers and potential outcomes:
  - If managed well, the AI boom could accelerate clean energy investment and foster more resilient supply chains.
  - If mismanaged, it risks locking in new emissions and deepening resource dependence.
  - Governments must integrate power plants, grids, water, and minerals into digital policy planning.

*Source: fd-dec25-lowres - references (December 2025).*

### 1.4 billion. It employs some 5,000 full-time staff mem-

### fd-dec25-lowres - 1.4 billion. It employs some 5,000 full-time staff mem-

### India: modernizing official statistics (interview with Saurabh Garg)
- Scope and staffing
  - Population reference: "1.4 billion."
  - Staffing: "some 5,000 full-time staff members at the central level and over 6,000 field investigators and supervisors during major survey operations across the nation’s 28 states and 8 territories."
- Technology and data collection
  - Door-to-door collection remains, but data collection is tablet based.
  - AI chatbots are being added to tablets to answer enumerators’ questions immediately.
  - All tablet data is uploaded directly to a background portal, called "e‑SIGMA," to make data processing easier.
  - Website improvements: mobile app, more infographics, direct unit-level data downloads, and training videos for researchers, students, policymakers, and stakeholders on accessing unit-level data.
- Incorporating alternative data and quality assurance
  - Exploring incorporation of "e-commerce, scanner, mobile, satellite, and other alternative data" into official statistics while emphasizing that official statistics "are underpinned by rigorous scientific methods and standards" and "go through comprehensive quality control to ensure accuracy, reliability, and comparability over time."
- Human and financial resources
  - Financial constraint: "We don’t have a budget constraint. We don’t need much money because our work is human- rather than finance-intensive."
  - Human resources structured in three levels:
    - Indian Statistical Service officers (top level) drawn from elite institutions.
    - Ground-level supervisors, many trained in statistics or mathematics.
    - Enumerators who are trained intensively.
  - Government-wide "Karmayogi" platform is encouraged for skill building (including communication).
- Labor statistics enhancements
  - Previously: annual labor force survey with quarterly updates covering only urban India.
  - Since January 2025: produced a "monthly report for both the urban and rural sectors."
  - Also have a "quarterly report" with more sectoral and employment detail.
  - Sample size: "We have nearly doubled the sample size." This increased data granularity.
  - New questionnaire content: employment status, education, training and skill levels, and where respondents graduated or received certification.
- Harmonization and metadata
  - Measures introduced to harmonize across agencies:
    - Registry of all government data sets, accounting for levels of importance.
    - A "national metadata structure" shared with all ministries.
    - Alignment with international systems of classification and national classifications to ensure internationally recognized definitions.
    - Use of unique identifiers for organizations and geographic locations, with each agency using its own identifier so that "two data sets can be read by each other."
    - A method for reconciling discrepancies between administrative data sets.
- Adoption of international standards and rebasing
  - India adopted the United Nations Fundamental Principles of Official Statistics in 2016 to underpin professional independence and accountability.
  - On SNA 2025 and rebasing:
    - "We’re in the process of rebasing our national accounts."
    - "We’re incorporating the changes required by SNA 2025 into our rebased methodology and new guidelines, which we expect to release over the next couple of years."
    - Recent sectoral accounts published annually; "this year, for instance, we are bringing out forest and water accounts."
    - Production-side robustness: the availability of GST [goods and services tax] data now provides more granular expenditure-side data for rebased GDP.
- Challenges highlighted
  - Need for more granular data at district, subdistrict, and village levels implies increasing sample sizes and increasing margin of error.
  - Frequency demands: users "aren’t willing to wait a year for data. Even a month is too long."
  - Ensuring cooperation and maintaining privacy and credibility in a social media age.

### Nowcasting and new data sources in developing economies
- Concept and origins
  - Nowcasting uses real-time and high-frequency data to predict the present, very recent past, and very near future of economic activity.
  - Originated in meteorology; in economics the term was popularized by Giannone, Reichlin, and Small (2008), with models developed from 2002 onward and adapted for macroeconomic nowcasting.
- Value and applications
  - Nowcasting provides timelier signals than official indicators during fast-moving events (for example, the pandemic).
  - Central banks and finance ministries use nowcasting to adjust forecasts, revise policy projections, and take timely actions.
  - Examples of high-frequency indicators: exports, imports, real-time consumer spending from electronic billing machines, wholesale and retail sales, livestock auctions, produce markets, satellite night-light intensity, Google search trends, tax authority electronic billing, and commercial agricultural data.
- Data gaps and frequency challenges
  - Official GDP publication frequency:
    - "Advanced economies and most emerging markets publish GDP quarterly. But about a third of countries in the world have only annual GDP."
    - First estimates of quarterly GDP in advanced economies typically appear about a month after quarter end—"two months in some major emerging market economies"—and in developing economies it can take "more than three months."
  - During crises, the delay in official indicators severely hampers policymaking.
- Case examples and findings
  - Rwanda
    - Launched a weekly economic activity index during COVID based on central bank measures and tax-authority electronic billing data.
    - Nowcasting included in staff briefings before quarterly Monetary Policy Committee meetings.
  - Kenya
    - Official quarterly GDP estimates usually released about "three months after a quarter ends."
    - Central bank uses nowcasting to start gauging the quarter "after only a week," using private consumer spending data; remittance data available after two weeks; trade, money supply, tourism, and electricity data in about 40 days to refine the picture.
  - South Africa
    - Commercial agricultural data gave an early read on food inflation; commercial indicators "lead CPI by one to three months across components."
  - Global research directions
    - Giannone’s latest research: using global and regional comovement to interpolate for countries with sparse data, exploiting GDP information from neighbors and trading partners.
    - Advances in large language models and AI create opportunities for exploiting text as data and integrating data with metadata.
- Limitations, cautions, and best practices
  - Nowcasting and alternative data cannot replace rigorous official data collection; they complement expanded official indicators.
  - Simple models often perform best; "start with the data"—identify high-quality or proxy data before selecting techniques.
  - New data sources can be noisy or nonrepresentative; models can fail when structural conditions change.
  - Broader concerns: privacy, transparency, legitimacy, governance, and capacity constraints (staffing, budgets, advanced computing).
  - Policy emphasis: invest in expanding official indicators, raise frequency and granularity, and build analytical capacity.

### Case studies and technological innovations (selected)
- Democratic Republic of the Congo (DRC)
  - Central bank nowcasting and forecasting and policy analysis system (IMF-assisted).
  - Combines traditional high-frequency indicators (copper and cobalt production and prices, money supply) with nontraditional inputs (satellite night-light intensity, Google search trends) to feed a quarterly projection model.
  - Contextual challenges: official GDP figures published only annually, high dollarization, mining sector contributes "about a third of GDP," and vulnerability to exchange rates and commodity prices.
- Guinea-Bissau (blockchain for wage bill)
  - May 2024 deployment (with IMF, EY, donors) of blockchain to manage public wage bill at ministries of finance and public administration.
  - Results: "The wage bill fell to half of tax revenue in 2024."
  - Planned expansion: platform will be expanded to cover "all 26,600 public officials and 8,100 pensioners nationwide."
- Kenya (real-time insights and nowcasting)
  - Nowcasting models can approximate Kenya’s growth ahead of official releases by exploiting comovement across indicators; Central Bank of Kenya and IMF cooperation on nowcasting and bimonthly surveys.
- Madagascar (AI customs)
  - October introduction of "agentic AI" to detect inconsistencies signaling fraud by cross-analyzing customs declarations, invoices, manifests, and internal/external databases.
  - Result: automated frontline tasks, allowing experts to focus on complex cases; builds a foundation for future AI-enabled tools.
- Broader finding: IMF technical assistance supports capacity development and adoption of these tools across countries.

### Policy recommendations, operational steps, and governance
- For national statistical systems
  - Make administrative and ministry data machine readable to permit collation, combination, and comparison of datasets.
  - Maintain a registry of government datasets and a national metadata structure; adopt internationally recognized definitions and classifications.
  - Use unique identifiers for organizations and geographic locations to enable interoperability.
  - Reconcile discrepancies among administrative datasets via formal methods.
  - Uphold the United Nations Fundamental Principles of Official Statistics to ensure professional independence, impartiality, and transparency.
  - Hold stakeholder consultations, publish detailed documentation for large-sample surveys, and organize conferences to engage users and gather technical insights.
- For data innovation and nowcasting
  - Prioritize "start with the data": identify high-quality domestic or proxy datasets before applying complex models.
  - Use alternative high-frequency data (billing, retail, satellite, search trends) to construct timely indicators, while recognizing limitations and representativeness issues.
  - Invest in analytical capacity: economists, statisticians, and data scientists, and expand computing resources.
  - Balance frequency and granularity with statistical reliability; acknowledge trade-offs between sample size, granularity, and margin of error.
- For governance and public trust
  - Ensure transparency, documentation, and independent scrutiny of methodologies.
  - Address privacy concerns and maintain public cooperation through clear safeguards.
  - Protect statistical agencies from political influence by adhering to established principles and ensuring professional independence.

*Source: https://www.imf.org/-/media/files/publications/fandd/article/2025/12/fd-dec25-lowres.pdf*

---


_Source: https://www.imf.org/-/media/files/publications/fandd/article/2025/12/fd-dec25-lowres.pdf_