## wpiea2026146

## Source details

**Canonical URL:** [wpiea2026146](https://www.imf.org/-/media/files/publications/wp/2026/english/wpiea2026146.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/wp/2026/english/wpiea2026146.pdf.md)
- [Structured JSON version](/-/media/files/publications/wp/2026/english/wpiea2026146.pdf.json)

---

### Background on crypto mining
- Blockchain plus a consensus protocol validates and records transactions; transaction records grouped into blocks cryptographically linked into a chain.
- Proof-of-work (PoW) mining (e.g., Bitcoin):
  - Miners compete to find a valid cryptographic hash; first to solve appends a block and is rewarded with newly issued coins plus transaction fees.
  - Puzzle difficulty increases with network participation, maintaining a relatively stable block creation rate.
- Hardware evolution and market structure:
  - Early mining used CPUs; competition and difficulty rendered general-purpose processors obsolete.
  - Major PoW mining now uses ASICs (application-specific integrated circuits) for higher hash rates and energy efficiency.
  - Mining pools aggregate individual miners’ computational power and split block rewards; as of January 2025, the top 10 pools accounted for approximately 95% of the Bitcoin network’s hash rate.

### Empirical strategy and data construction
- Rationale: Mining activity is not directly observable; ASIC trade flows in customs data provide an observable proxy for geographic distribution of mining hardware.
- Data sources and HS codes:
  - Primary data: 8-digit Harmonized System (HS) export data from the Chinese mainland, disaggregated by destination country and product code.
  - Candidate HS8 codes: 8471.41.90, 8471.50.40, 8473.30.90.
    - Baseline focuses on 8471.41.90 and 8471.50.40; 8473.30.90 used in robustness checks (quantity measure often in kilograms and lower unit values).
  - All major ASIC producers located in the Chinese mainland.
- Hong Kong SAR handling:
  - Approximately 32 percent of the Chinese mainland’s reported ASIC exports are shipped to Hong Kong SAR.
  - Exports from the Chinese mainland to Hong Kong SAR are excluded from the bilateral dataset (Hong Kong SAR not treated as a final destination).
  - Hong Kong SAR re-exports (UN COMTRADE) are incorporated on the same footing as exports originating from the Chinese mainland.
  - Manual mapping of HS codes between Chinese mainland and Hong Kong SAR using official Hong Kong documentation.
- Aggregation and indices:
  - Monthly panel of mining-hardware imports constructed for January 2018 to August 2023.
  - Two aggregate indicators:
    - Value index: sum of dollar value of relevant HS codes per country.
    - Laspeyres-type volume index (base year 2018; average value in 2018 = 100):
      - I^q_j(t) = sum_i s^2018_{i,j} q_{i,j}(t) / q̄^2018_{i,j} × 100,
        where i indexes exporter-HS8 pairs, s^2018_{i,j} is 2018 expenditure share, q̄^2018_{i,j} is average monthly import quantity.
  - GPUs collected from supplementary trade sources (including Taiwan Province of China) used only for robustness checks because GPUs are multipurpose.

### Stylized facts validating the proxy
- Fact 1: Correlation with independent hashrate estimates
  - Strong correlation between import-based proxy and CCAF country-level hashrate estimates (September 2019–January 2022), despite CCAF assumptions possibly fragile where mining is prohibited or heavily regulated.
  - Customs-based records drawn from hardware-exporting countries are harder to mask; strong correlation provides validation for the proxy.
- Fact 2: Global surge in 2021–2022 aligned with crypto prices
  - Three phases observed (2018–Aug 2023):
    1. 2018–20 (stable and deflationary): volume hovered around 2018 average; value index drifted down.
    2. 2021–early 2022 (boom): 12-month moving average of volume about 25% above baseline by early 2022; import values rebounded strongly.
    3. Late 2022–23 (retrenchment): indices trended back toward pre-pandemic averages.
  - ASIC exports exhibited a sharper surge and decline in 2022–2023 mirroring Bitcoin price movements—distinguishing crypto-specific dynamics from general post-pandemic supply-chain normalization.
- Fact 3: Cross-country heterogeneity
  - Annual growth rates vary widely across countries and time; synchronized distribution shifts indicate global drivers (crypto price cycles).
  - Example statistics:
    - In 2021 many countries had a median growth rate of 43 percent in mining hardware imports.
    - Argentina experienced roughly 700 percent higher annual mining hardware imports than the 2018–2020 average during the 2021–2022 surge.

### Theoretical mapping of risky returns and comparative statics
- Returns definitions:
  - Domestic asset return (in foreign currency): μd = E[1 + r S1/S0].
  - Bitcoin-mining hardware return (per normalized unit 1 ASIC → 1 bitcoin):
    - μb = E[(Pb − Pe e(T)) / Pa], where Pa, Pb, Pe denominated in foreign currency; e(T) captures electricity consumption as an increasing function of temperature T.
  - Pa and Pe treated as known at investment decision time; risk comes from future Pb.
- Calibration baseline (used in Figure 5):
  - μf = 0, μd = 1, σd^2 = 1, E(Pb) = 3, σb^2 = 3, Pe = 1.5, T = 20, Pa = 1, γ = 2, ρ = 2, e(T) = 0.43 + 0.0206T + 0.0004T^2.
- Comparative-statics conclusions:
  - Elevated domestic uncertainty and suppressed domestic returns encourage crypto mining.
  - High expected Bitcoin price, low perceived Bitcoin risk, cheap hardware increase attractiveness.
  - Low electricity prices and cooler climates lower operating costs and favor mining.
  - Stronger capital controls (higher γ) expand mining as a way around financial-account restrictions.

### Empirical identification of surges and regression framework
- Surge detection (monthly series m_t = log volume of mining-hardware imports):
  - For candidate surge start τ between January 2020 and May 2023:
    1. Fit linear trend on m_t using January 2018 to τ−1; RMSE_τ computed.
    2. Predicted value ˆm_{τ,t}; normalized deviation ε_{τ,t} = (m_t − max{ˆm_{τ,t}, 0}) / RMSE_τ for t ≥ τ.
    3. Average post-τ excess imports E_τ = (1/(T − τ + 1)) Σ_{t=τ}^T ε_{τ,t}, with T = August 2023.
    4. Choose τ* that maximizes E_τ; set ε_{τ*,t} = 0 for t < τ* and treat subsequent ε as the country surge episode.
  - Search begins January 2020 to use pre-pandemic baseline (2018–2019); procedure identifies surge onset but not end dates.
- Regression specification (equation (4)):
  - ε_{j,t} = a0 + a1 ln Pb,t + a2 σb,t + a3 ln Pa,t + a4 ln T_j + a5 ln Pe,j,t + a6 μd,j,t + a7 σd,j,t + a8 γ_j.
  - Expected signs from model: a1, a7, a8 positive; remaining coefficients negative.
- Variable construction:
  - Quarterly frequency (daily/monthly inputs averaged to quarter).
  - Pb,t from Haver Analytics; σb,t = standard deviation of daily returns within each quarter.
  - T_j from Redivis; Pe,j,t from Global Petrol Prices.
  - μd,j,t measured by UIP wedge: μd,j,t = r_{j,t} − (s^e_{j,t+1} − s_{j,t}) − r^*_t; exchange rates from IFS; 12-month-ahead expectations from Consensus Economics.
  - σd,j,t = 12-month backward-looking volatility of UIP wedge.
  - γ_j = capital control indices of Fernández et al. (2016), last-available 2019 value.
  - Appendix lists 45 countries included in estimation.

### Effective ASIC price index construction
- Steps:
  1. Compile public ASIC prices by energy-efficiency tiers.
  2. Weight prices by efficiency-adjusted performance; penalize inefficient hardware.
  3. Scale index by evolving Bitcoin-network difficulty.
- Quality-adjustment detail:
  - Price P and efficiency E (J/TH) transformed so effective quality-unit cost = P * E; uses midpoints of five efficiency tiers from Hashrate Index and unweighted mean of quality-adjusted prices.
  - Difficulty scaling uses Bitcoin network difficulty from Blockchain.com.
- Block reward unchanged between May 2020 and April 2024 for nearly all identified surges.

### Regression results, robustness, and key quantitative estimates
- General empirical findings:
  - Global drivers:
    - ln(Pb,t) (Bitcoin price): positive and highly significant across specifications.
    - σb,t (Bitcoin return volatility): negative and highly significant.
    - ln(Pa,t) (effective ASIC price): negative and significant.
  - Domestic operating conditions:
    - ln(T_j) (average temperature) and ln(Pe,j,t) (electricity price) negative; coefficients negative and significant in narrow specifications.
  - Domestic returns/volatility:
    - μd,j,t has expected negative sign but is statistically insignificant.
    - σd,j,t shows limited and sometimes counterintuitive evidence (marginal significance under bootstrap at 10%).
  - Capital controls:
    - Coefficient positive as predicted but statistically insignificant in baseline; similar results when replacing F_kao_j with broader F_ka_j.
- Robustness:
  - Expanding HS code set from narrow (ASICs) to medium and broad (including GPUs) preserves global factor signs; effect sizes decline and domestic variables lose significance.
  - Aggregating customs-union members (Mercosur, EU) into synthetic units yields coefficient estimates and significance close to baseline.
  - GPUs used in robustness: GPUs comprised between 20 and 50 percent of value of crypto mining-related exports and share rose during sample; GPU exports largely from Hong Kong SAR, Chinese mainland, and increasingly Taiwan Province of China.
- Key coefficient estimates (selected, from Table 2; standard errors in parentheses):
  - Sample: Observations = 725; R-squared across specs: 0.265, 0.265, 0.266, 0.209, 0.230.
  - ln(Pb,t):
    - (1) Narrow: 0.837*** (0.148)
    - (2) Narrow (Bootstrap): 0.837*** (0.151)
    - (3) Narrow (Clustered): 0.834*** (0.147)
    - (4) Medium: 0.663*** (0.115)
    - (5) Broad: 0.647*** (0.105)
  - σb,t:
    - (1) −0.977*** (0.264)
    - (2) −0.977*** (0.272)
    - (3) −0.980*** (0.263)
    - (4) −0.787*** (0.182)
    - (5) −0.804*** (0.180)
  - ln(Pa,t):
    - (1) −0.467*** (0.148)
    - (2) −0.467*** (0.144)
    - (3) −0.465*** (0.148)
    - (4) −0.281** (0.120)
    - (5) −0.249** (0.105)
  - ln(T_j):
    - (1) −1.846*** (0.573)
    - (2) −1.846*** (0.588)
    - (3) −1.869*** (0.576)
    - (4) −0.547 (0.634)
    - (5) −0.787* (0.438)
  - ln(Pe,j,t):
    - (1) −3.393* (1.878)
    - (2) −3.393* (2.025)
    - (3) −3.347* (1.881)
    - (4) −3.108 (2.119)
    - (5) −2.803 (2.046)
  - μd,j,t:
    - (1) −0.021 (0.035)
    - (2) −0.021 (0.033)
    - (3) −0.021 (0.034)
    - (4) −0.040 (0.032)
    - (5) −0.031 (0.031)
  - σd,j,t:
    - (1) −0.211 (0.137)
    - (2) −0.211* (0.121)
    - (3) −0.218 (0.139)
    - (4) 0.041 (0.090)
    - (5) 0.020 (0.086)
  - F_kao_j (capital outflow control index):
    - (1) 0.264 (0.304)
    - (2) 0.264 (0.344)
    - (3) 0.093 (0.371)
  - F_ka_j in column (3): 0.333 (0.371)
  - Constants range ~10.892** (5.336) to 6.635* (3.886) across specs.
- Robustness to customs-union re-exports (Table 4, selected estimates):
  - ln(P_b,t): 0.804*** (0.152); 0.681*** (0.169); 0.697*** (0.190) for Mercosur, EU, and both aggregations.
  - σ_b,t: −0.965*** (0.281); −0.787*** (0.283); −0.787** (0.302).
  - ln(T_j): −1.834*** (0.592); −2.073*** (0.709); −2.167*** (0.742).
  - Observations: 699; 562; 511. R-squared: 0.249; 0.286; 0.272.
- Illustrative magnitudes (Table 3 discussion):
  - Global time-series shift (2020 → 2021 averages):
    - Bitcoin price: 10,931 → 46,959; Implied effect: 1.22
    - Bitcoin return volatility: 4.04 → 3.87; Implied effect: 0.17
    - ASIC price index (2020 avg=100): 100 → 394.69; Implied effect: −0.64
    - Total Panel A effect: 0.75
    - Memo: Median excess imports: 0.40 → 1.16; Implied effect: 0.76
  - Cross-sectional 10th→90th percentile shifts (Panel B):
    - Temperature (°C): 9.62 → 26.60; Implied effect: 0.84
    - Electricity price (USD/kWh): 0.07 → 0.21; Implied effect: 0.4
    - Domestic excess return in FX (%): −0.29 → 4.73; Implied effect: 0.11
    - Capital controls on outflows: 0 → 0.95; Implied effect: 0.25
    - Total Panel B effect: 1.60
    - Memo: Average excess imports: 0.03 → 1.95; Implied effect: 1.92

### Additional empirical observations and data highlights
- ASIC exports vs aggregate Chinese exports: ASIC exports rise later and more sharply and display a boom–bust cycle closely mirroring Bitcoin price, unlike aggregate exports.
- GPUs and market context:
  - GPUs relevant for Ethereum mining prior to September 15, 2022 (Ethereum PoW → PoS transition).
  - GPUs comprised between 20 and 50 percent of value of crypto mining-related exports; share rose during sample.
  - GPU export sources: Hong Kong SAR, Chinese mainland, rising exports from Taiwan Province of China.
- Cross-country magnitudes:
  - Mining-related imports represent a significant fraction of GDP for many small economies and trade hubs (examples cited: Marshall Islands, Bhutan, Singapore, the Netherlands).
  - Top increases in mining-related imports as share of GDP from 2019 to 2021 primarily in developing countries; some cases rose from near zero to large GDP shares.

### Policy implications and suggested directions
- Domestic policy effects:
  - Untargeted subsidies that keep electricity prices artificially low can encourage excessive mining activity, exacerbating financial distortions and environmental costs.
- Macro and regulatory interactions:
  - Capital controls and weak domestic investment opportunities may render mining an outlet for capital flight; policymakers should consider interactions between capital-account measures and crypto mining incentives.
- Suggested future research:
  - Apply the framework to study regulatory changes, mining bans, and links between mining intensity and illicit financial flows.

*Source: wpiea2026146 - selected extracts from sections 2.1, 2.2, 3.3, Table 1, Table 3, Table 4, and appendices (as provided).*

### 2.1    Background on Crypto Mining

### 2.1    Background on Crypto Mining

### Blockchain and proof-of-work mining
- Blockchain, together with a consensus protocol, underpins decentralized validation and recording of transactions in crypto asset networks. Transaction records are grouped into blocks cryptographically linked to form a continuously growing chain.
- In blockchains using the proof-of-work (PoW) consensus protocol, such as Bitcoin, specialized nodes called miners compete to solve a computationally intensive puzzle—finding a valid cryptographic hash.
- The first to find a valid solution earns the right to append a new block to the chain and is rewarded with newly issued coins. This mechanism ensures consensus and security and represents the sole channel through which new crypto assets are created.
- Solving the puzzle requires substantial computational power and becomes progressively more difficult as network participation grows, maintaining a relatively stable rate of block creation.
- Mining’s total financial incentive also includes transaction fees paid by users from already circulating assets.

### Mining hardware evolution and market structure
- Early Bitcoin mining could be performed using standard CPUs; growing competition and algorithmic difficulty rendered general-purpose processors obsolete.
- Today, virtually all mining for major PoW-based cryptocurrencies is conducted using application-specific integrated circuits (ASICs)—custom-built chips that offer vastly higher hash rates and energy efficiency.
- As individual mining became less profitable, most activity migrated to mining pools, where participants combine computational power and receive predictable income (in bitcoin terms) by splitting block rewards according to contribution.
- As of January 2025, the top 10 pools accounted for approximately 95% of the Bitcoin network’s hash rate.

### Empirical strategy: mining hardware as an observable proxy
- Mining activity is not directly observable, but the specialized hardware required (ASICs) leaves detectable traces via trade flows.
- Analogy: an observer unaware of where a historical gold rush occurred could infer the location by tracing inputs (pickaxes, sluices) rather than the gold itself.
- The empirical approach exploits the specificity of currently used mining hardware to infer geographic distribution of mining from customs records.

### Identifying mining hardware imports in customs data
- All major producers of crypto mining ASICs are located in the Chinese mainland.
- Data source: 8-digit Harmonized System (HS) export data from China’s customs authority, disaggregated by destination country and product code.
- Candidate HS8 codes for ASIC exports identified: 8471.41.90, 8471.50.40, and 8473.30.90.
  - 8473.30.90 likely corresponds to ASIC components or accessories (evidenced by significantly lower unit values); baseline analysis focuses on 8471.41.90 and 8471.50.40, with robustness checks including the broader set.
- Approximately 32 percent of the Chinese mainland’s reported ASIC exports are shipped to Hong Kong SAR.
  - Given Hong Kong SAR’s role as a logistics and re-export hub and the lack of evidence of significant local crypto mining, exports from the Chinese mainland to Hong Kong SAR are excluded from the bilateral dataset; Hong Kong SAR is not treated as a final destination.
  - Hong Kong SAR’s re-exports of the relevant HS codes (as reported in UN COMTRADE) are incorporated on the same footing as exports originating from the Chinese mainland.
  - Because Hong Kong SAR uses slightly different HS codes than the Chinese mainland, product codes are manually mapped across the two systems using official documentation from the Hong Kong SAR Census and Statistics Department.
- Outside the Hong Kong SAR special case, recorded export destinations are treated as the best available proxy for the location where mining hardware is ultimately used.
- GPUs (graphics processing units) data from supplementary trade sources (including Taiwan Province of China) are collected for robustness analysis; GPUs are multipurpose and thus used only for robustness checks.

### HS code correspondences (Chinese mainland ↔ Hong Kong SAR)
- 8471.41.90 ↔ 8471.41.00 — Other automatic data processing machines; comprising in the same housing at least a CPU and an input and output unit, including combined
- 8471.50.40 ↔ 8471.50.00 — Processing units other than those of subheading 8471.41 and 8471.49, whether or not containing in the same housing one or two of the following types of units: storage units, input units, output units
- 8473.30.90 ↔ 8473.30.90 — Parts and accessories for automatic data processing machines and units thereof, magnetic or optical readers, transcribing machines, etc., nesoi

### Aggregation and index construction
- For each country, monthly imports under the identified HS8 codes are aggregated into a single “synthetic” product capturing overall import of mining hardware.
- Two aggregate indicators developed:
  - Value index: sum of the dollar value of all relevant HS codes for each country (conflates quantities and prices).
  - Laspeyres-type volume index to capture changes in quantity of mining-related hardware imports:
    - Trade data provide quantities either in number of units or in kilograms.
    - Each exporter-HS8 code pair treated as a separate product; year 2018 used as the base period.
    - Monthly import volume index I^q_j(t) given by:
      I^q_j(t) = sum_i s^2018_{i,j} q_{i,j}(t) / q̄^2018_{i,j} × 100,
      where i indexes distinct exporter-HS8 code pairs, s^2018_{i,j} is the expenditure share of product i in 2018, and q̄^2018_{i,j} is the average monthly import quantity of product i for country j.
    - The average value of the index in 2018 is 100.
- Aggregated data construct a monthly panel dataset of mining hardware imports from January 2018 to August 2023.
- Note: The HS8 code with a quantity measure in kilograms is 8473.30.90, used only in robustness checks.

### Stylized facts validating the proxy and motivating analysis
- Fact 1: Import shares are correlated with independent estimates of country-level hashrates.
  - Strong correlation observed between the import-based proxy and country-level hashrate estimates from the Cambridge Centre for Alternative Finance (CCAF), which monitored Bitcoin mining activity between September 2019 and January 2022 using data from multiple mining pools.
  - CCAF estimates rely on two assumptions: (i) IP addresses reported to mining pools accurately reflect miners’ physical locations, and (ii) membership of mining pools included in the sample is representative of Bitcoin’s global hashrate distribution. These assumptions may be fragile where mining is prohibited or heavily regulated.
  - Customs-based approach draws on hardware-exporting countries’ records, which are more difficult to mask or manipulate. While re-exports or final country of use cannot be detected perfectly, the strong correlation with CCAF estimates provides validation for the proxy.
- Fact 2: Global imports of mining hardware surged in 2021–2022, coinciding with large swings in crypto assets prices.
  - Three distinct phases in sample period:
    1. 2018-20 (stable and deflationary): Volume hovered around its 2018 average; value index drifted down with declining unit costs due to technological improvements.
    2. 2021–early 2022 (boom): Volume surged, with the 12-month moving average climbing to about 25% above its baseline by early 2022. Import values rebounded more strongly, reflecting rising hardware prices as demand outpaced efficiency gains amid the crypto rally.
    3. Late 2022–23 (retrenchment): Both indices trended back toward their pre-pandemic average.
  - Timing: Surge in 2021–2022 aligns with rapidly rising crypto asset prices; subsequent retrenchment coincides with the decline in crypto asset prices. The comovement is consistent with investment in mining hardware responding to crypto market conditions.
  - Alternative interpretation: Surge could reflect normalization of global supply chains post-COVID-19. Appendix analyses compare ASIC-related exports with aggregate Chinese exports and Bitcoin price; ASIC exports show a markedly different, sharper surge followed by pronounced decline in 2022–2023, mirroring Bitcoin price movements—suggesting crypto-specific investment dynamics beyond a generic post-pandemic catch-up.
- Fact 3: Import growth rates vary widely across countries and over time.
  - Annual growth rates of mining hardware imports show substantial cross-country variation, indicating importance of local factors.
  - Synchronized shifts in distribution of growth rates over time point to global factors (e.g., crypto price cycles).
  - Example: In 2021 many countries experienced a boom in imports of crypto mining equipment, with a median growth rate of 43 percent.
  - Country example: Argentina experienced roughly 700 percent higher annual mining hardware imports than the 2018-2020 average during the 2021–2022 surge; temporal alignment with CCAF’s hashrate estimates reinforces the trade-based measure’s validity as a proxy for mining investment.

*Italic: Source: wpiea2026146 - 2.1    Background on Crypto Mining (extracted from the supplied IMF content).*

### 3.3    Unpacking Risky Returns and Comparative Statics

### 3.3    Unpacking Risky Returns and Comparative Statics

### Theoretical mapping of risky returns
- Domestic asset return (in foreign currency):
  - μd = E[1 + r S1/S0].
  - If the domestic interest rate is risk-free in local currency, all risk comes from unexpected exchange rate movements; randomness can also capture general domestic financial-market uncertainty.
- Bitcoin-mining hardware return (per normalized unit where 1 ASIC → 1 bitcoin):
  - μb = E[(Pb − Pe e(T)) / Pa],
    - where Pa, Pb, and Pe are prices of ASICs, Bitcoin, and electricity, each denominated in foreign currency,
    - e(T) is a positive-valued increasing function of temperature T capturing electricity consumption for operating ASICs.
  - Pa and Pe can be treated as known at the investment decision time; risk comes from future Pb.
- Assumption rationales:
  - Bitcoin price is determined on the global market, implying returns on the domestic asset and Bitcoin mining are uncorrelated.
  - Normalization 1 ASIC = 1 bitcoin requires adjusting effective ASIC prices for changes in Bitcoin network difficulty (see section 4.2).

### Comparative statics (numerical illustration summarized)
- Calibration baseline used in Figure 5:
  - μf = 0, μd = 1, σd^2 = 1, E(Pb) = 3, σb^2 = 3, Pe = 1.5, T = 20, Pa = 1, γ = 2, ρ = 2, e(T) = 0.43 + 0.0206T + 0.0004T^2.
- Comparative-statics conclusions:
  - Elevated domestic uncertainty and suppressed domestic returns encourage crypto mining.
  - High expected Bitcoin price with low perceived Bitcoin risk and cheap hardware increase mining attractiveness.
  - Low electricity prices and cooler climates favor mining by lowering e(T) and operating costs.
  - Stronger capital controls (higher γ) expand mining as a way around financial account restrictions.

### Empirical strategy: identifying surges in mining-hardware imports
- Define m_t = log volume of country’s mining-hardware imports (monthly), using the synthetic HS8 product index.
- For each candidate surge start month τ between January 2020 and May 2023:
  1. Estimate a linear time trend on m_t using January 2018 to τ−1; denote RMSE_τ as the trend RMSE.
  2. Predicted value for month t ≥ τ is ˆm_{τ,t}; normalized deviation (excess imports):
     - ε_{τ,t} = (m_t − max{ˆm_{τ,t}, 0}) / RMSE_τ.
  3. Average post-τ excess imports:
     - E_τ = (1/(T − τ + 1)) Σ_{t=τ}^T ε_{τ,t}, with T = final month (August 2023).
  4. Choose τ* that maximizes E_τ; set ε_{τ*,t} = 0 for t < τ* and treat subsequent excess imports as the country’s surge episode.
- Rationale and implementation notes:
  - Normalization by RMSE adjusts for baseline differences and reduces influence of persistent level differences across countries.
  - Search begins January 2020 with pre-pandemic baseline (2018–2019) because of the COVID-19 crypto boom and China’s regulatory crackdown.
  - The procedure does not attempt to identify surge end dates; all deviations after onset are included.

### Regression specification connecting model to data
- Interpreting excess imports ε_{j,t} as an increase in country j’s ASIC-share x_b motivates:
  - ε_{j,t} = a0 + a1 ln Pb,t + a2 σb,t + a3 ln Pa,t + a4 ln T_j + a5 ln Pe,j,t + a6 μd,j,t + a7 σd,j,t + a8 γ_j.  (equation (4))
- Coefficient interpretations and model predictions:
  - a1, a2, a3 capture global profitability drivers: Bitcoin price, Bitcoin return volatility, and effective ASIC price.
  - a4, a5 represent operating-cost determinants: average temperature and electricity prices.
  - a6, a7 reflect domestic asset expected return and risk.
  - a8 captures capital-account frictions (capital controls).
  - Expected signs from the model: a1, a7, and a8 positive; remaining coefficients negative.
- Variable construction details:
  - Quarterly frequency (averaging daily/monthly inputs).
  - Pb,t from Haver Analytics; σb,t = standard deviation of daily returns within each quarter.
  - T_j from Redivis; Pe,j,t from Global Petrol Prices.
  - μd,j,t (expected excess return on domestic assets in foreign currency) measured by UIP wedge:
    - μd,j,t = r_{j,t} − (s^e_{j,t+1} − s_{j,t}) − r^*_t, where r_{j,t} and r^*_t are domestic and US interest rates; s_{j,t} and s^e_{j,t+1} are spot and expected log exchange rate (local currency per USD).
    - Exchange rates from IFS; 12-month-ahead exchange-rate expectations from Consensus Economics survey forecasts.
  - σd,j,t measured using 12-month backward-looking volatility of the UIP wedge.
  - γ_j measured by capital control indices of Fernández et al. (2016), using last-available 2019 value.
  - Appendix lists 45 countries included in the estimation.

### Effective ASIC Price Index (construction summary)
- Steps to construct a model-consistent effective ASIC price index:
  1. Compile public data on ASIC prices by energy-efficiency tiers.
  2. Weight prices by efficiency-adjusted performance, penalizing inefficient hardware.
  3. Scale the index by evolving Bitcoin-network difficulty.
- Result: a time series of ASIC prices reflecting the true hardware cost of generating new bitcoins; full details in Appendix A.3.

### Regression results and robustness
- General findings:
  - Global drivers: an increase in Bitcoin price is associated with higher mining-hardware excess imports; greater Bitcoin return volatility reduces mining investment; higher ASIC prices lower excess imports.
  - Domestic operating conditions: higher average temperature and higher energy prices reduce mining-hardware excess imports (negative and significant coefficients).
  - Domestic returns and volatility: limited statistical evidence—expected domestic return has expected sign but is insignificant; domestic-return volatility has counterintuitive sign and is marginally significant only under bootstrap at the 10% level.
  - Capital controls: coefficient positive as predicted but statistically insignificant; similar results when replacing outflow index (kao) with broader index (ka).
- Robustness checks:
  - Results robust to expanding HS code set from narrow (ASICs) to medium and broad (including GPUs); global factors retain expected signs and significance, though effect sizes decline and domestic variables lose significance.
  - Two empirical-design strategies mitigate HS-code measurement concerns:
    1. Focus on surges (deviations from trend) rather than levels.
    2. Re-estimate using broader HS sets including GPUs.

### Key quantitative estimates (selected coefficients from Table 2)
- Sample: Observations = 725; R-squared varies across specifications: 0.265, 0.265, 0.266, 0.209, 0.230.
- Coefficients (standard errors in parentheses):
  - ln(Pb,t):
    - Column (1) Narrow: 0.837*** (0.148)
    - Column (2) Narrow (Bootstrap): 0.837*** (0.151)
    - Column (3) Narrow (Clustered): 0.834*** (0.147)
    - Column (4) Medium: 0.663*** (0.115)
    - Column (5) Broad: 0.647*** (0.105)
  - σb,t:
    - (1) −0.977*** (0.264)
    - (2) −0.977*** (0.272)
    - (3) −0.980*** (0.263)
    - (4) −0.787*** (0.182)
    - (5) −0.804*** (0.180)
  - ln(Pa,t):
    - (1) −0.467*** (0.148)
    - (2) −0.467*** (0.144)
    - (3) −0.465*** (0.148)
    - (4) −0.281** (0.120)
    - (5) −0.249** (0.105)
  - ln(T_j):
    - (1) −1.846*** (0.573)
    - (2) −1.846*** (0.588)
    - (3) −1.869*** (0.576)
    - (4) −0.547 (0.634)
    - (5) −0.787* (0.438)
  - ln(Pe,j,t):
    - (1) −3.393* (1.878)
    - (2) −3.393* (2.025)
    - (3) −3.347* (1.881)
    - (4) −3.108 (2.119)
    - (5) −2.803 (2.046)
  - μd,j,t:
    - (1) −0.021 (0.035)
    - (2) −0.021 (0.033)
    - (3) −0.021 (0.034)
    - (4) −0.040 (0.032)
    - (5) −0.031 (0.031)
  - σd,j,t:
    - (1) −0.211 (0.137)
    - (2) −0.211* (0.121)
    - (3) −0.218 (0.139)
    - (4) 0.041 (0.090)
    - (5) 0.020 (0.086)
  - F_kao_j (capital outflow control index):
    - (1) 0.264 (0.304)
    - (2) 0.264 (0.344)
    - (3) 0.093 (0.371)
    - (4) −0.032 (0.354)
  - F_ka_j (broader capital control index) in column (3): 0.333 (0.371)
  - Constant terms range from 10.892** (5.336) in narrow specs to 6.635* (3.886) in broad spec.
- In-sample illustrative magnitudes (from Panels A and B, Table 3 discussion):
  - Global market conditions moving from 2020 average to 2021 average (Bitcoin and ASIC prices roughly quadrupled; volatility slightly decreased) imply a 0.75 increase in the excess imports measure—close to the median country’s observed change.
  - Cross-country decomposition: a hypothetical country with the most favorable mining conditions (cool climate, cheap electricity, low domestic returns, stringent capital controls) would have excess imports 1.6 units higher, on average, than a country with the least favorable mix.

*wpiea2026146 - 3.3    Unpacking Risky Returns and Comparative Statics (IMF Working Paper).*

### 2.2 and Table 1); “Broad” also includes GPU-related codes (see Appendix A.4).  Sample:  2019Q1-2023Q1.

### 2.2 and Table 1); “Broad” also includes GPU-related codes (see Appendix A.4). Sample: 2019Q1-2023Q1.

### Illustrative decomposition of excess imports (Table 3: Global and local factors)
- Panel A — Global factors (time series): 2020 vs 2021 average values; implied partial effects (based on estimates from Table 2, column (1))
  - Bitcoin price (USD/BTC): 10,931 → 46,959; Implied effect: 1.22
  - Bitcoin return volatility: 4.04 → 3.87; Implied effect: 0.17
  - ASIC price index (2020 avg=100): 100 → 394.69; Implied effect: -0.64
  - Total effect (Panel A): 0.75
  - Memo: Median excess imports: 0.40 → 1.16; Implied effect: 0.76
- Panel B — Country-specific factors (cross-section): 10th vs 90th percentile values (each variable averaged over 2019Q1–2023Q1); direction chosen so implied effect is positive
  - Temperature (°C): 9.62 → 26.60; Implied effect: 0.84
  - Electricity price (USD/kWh): 0.07 → 0.21; Implied effect: 0.4
  - Domestic excess return in FX (%): -0.29 → 4.73; Implied effect: 0.11
  - Capital controls on outflows: 0 → 0.95; Implied effect: 0.25
  - Total effect (Panel B): 1.60
  - Memo: Average excess imports: 0.03 → 1.95; Implied effect: 1.92
- Note: Table 3 uses partial effects from shifting variables (Panel A time-series shifts; Panel B cross-sectional 10th→90th percentile shifts) and compares hypothetical countries with most and least favorable mixes for mining.

### Regression robustness to customs-union re-exports (Table 4 summary)
- Purpose: Assess sensitivity of baseline regressions to potential re-exports within customs unions by aggregating members into synthetic units (Mercosur, EU, and both).
- Specification: Estimating equation (4) using the narrow ASIC-related HS 8-digit product group. Sample: 2019Q1–2023Q1.
- Key coefficient estimates (columns: (1) Mercosur, (2) EU, (3) Mercosur and EU), standard errors in parentheses; significance levels: *** p<0.01, ** p<0.05, * p<0.1
  - ln(P_b,t): 0.804*** (0.152); 0.681*** (0.169); 0.697*** (0.190)
  - σ_b,t: -0.965*** (0.281); -0.787*** (0.283); -0.787** (0.302)
  - ln(P_a,t): -0.447*** (0.150); -0.299* (0.176); -0.348* (0.189)
  - ln(T_j): -1.834*** (0.592); -2.073*** (0.709); -2.167*** (0.742)
  - ln(P_e,j,t): -3.212* (1.821); -6.942*** (2.175); -6.320** (2.498)
  - μ_d,j,t: -0.020 (0.038); -0.022 (0.038); -0.027 (0.041)
  - σ_d,j,t: -0.212 (0.156); -0.220 (0.149); -0.236 (0.169)
  - F_kao_j: 0.357 (0.316); 0.058 (0.345); 0.128 (0.364)
  - Constant: 10.692* (5.989); 11.380* (6.119); 12.487* (6.767)
  - Observations: 699; 562; 511
  - R-squared: 0.249; 0.286; 0.272
- Interpretation: Coefficient estimates and significance remain close to baseline results, indicating robustness to aggregation for customs unions.

### Main empirical findings and interpretations
- Global drivers strongly influence timing and magnitude of mining investment:
  - Crypto asset prices (Bitcoin price) and ASIC hardware costs are primary global determinants of mining hardware import surges.
- Local fundamentals and policies modulate intensity and location of mining activity:
  - Cooler temperatures and cheaper electricity correlate with more pronounced mining activity across countries.
  - Suggestive evidence that mining may be more attractive where domestic investment opportunities are weaker and capital flows are more restricted (consistent with mining serving as a form of capital flight under stringent capital controls).
- Trade-data interpretation and robustness:
  - Broader IT hardware categories (e.g., GPUs) likely attenuate mining signal due to alternative uses of multi-purpose hardware.
  - Results robust to allowing for potential re-exports within customs unions (Mercosur and EU aggregation).
  - Baseline approach assumes recorded export destinations correspond to countries of ultimate operation; robustness checks aggregating customs-union members produce similar estimates.

### Methodology and data highlights
- Novel measurement approach:
  - Uses publicly available customs data from dominant exporters of crypto mining hardware to construct country-level indices of mining hardware imports and identify surges (sharp increases in hardware inflows proxying for large-scale investment in mining capacity).
  - Provides a transparent, globally consistent alternative to proprietary IP-based tracking.
- ASIC price index construction (model-consistent):
  - Quality adjustment: price P and efficiency E (J/TH) transformed so effective quality-unit cost = P * E; uses midpoints of five efficiency tiers from Hashrate Index and unweighted mean of quality-adjusted prices.
  - Difficulty scaling: multiply quality-adjusted index by Bitcoin network difficulty (from Blockchain.com) to obtain final effective ASIC price index.
  - Block reward unchanged between May 2020 and April 2024 for nearly all identified surges, allowing this factor to be ignored.
- GPU tracing (Appendix A.4):
  - GPUs relevant for Ethereum mining prior to September 15, 2022 (Ethereum PoW → PoS transition).
  - GPU export data obtained for sample period largely preceding AI demand surge (ChatGPT released November 2022), reducing concerns about non-mining drivers.
  - Sources and HS codes used:
    - Taiwan Province of China: HS code 8471.80.00 (customs agency)
    - Chinese mainland: HS code 8471.80.00 (customs agency)
    - Hong Kong SAR: re-exports under HS code 8471.80.90 (UN COMTRADE)
  - Industry context: about 7 percent share of total crypto assets market capitalization for Ethereum as of May 2025; downstream assembly/testing concentrated in East Asia (Deloitte: ~three-quarters of revenue in Taiwan Province of China).
- Sample and country coverage:
  - Empirical panel of 45 countries (listed in Appendix A.1), sample period 2019Q1–2023Q1 (for primary analyses).

### Additional empirical observations (from appendices and figures)
- ASIC export dynamics vs aggregate Chinese exports:
  - ASIC exports rise later and more sharply than aggregate exports during post-pandemic recovery and exhibit a boom–bust cycle closely mirroring Bitcoin price movements (contrasting with steadily elevated aggregate exports).
- GPU share and exports:
  - GPUs comprised between 20 and 50 percent of the value of all crypto mining-related exports (ASICs and GPUs) and the share was increasing during the sample period.
  - GPU exports largely sourced from Hong Kong SAR and the Chinese mainland, with a recent significant rise in Taiwan Province of China exports.
- Cross-country patterns and magnitudes:
  - Figure-based evidence: for many small economies and trade hubs (e.g., Marshall Islands, Bhutan, Singapore, the Netherlands) mining-related imports represent a significant fraction of GDP.
  - Top increases in mining-related imports as share of GDP from 2019 to 2021 are dominated by developing countries; some cases show growth from near zero to large shares of GDP (example cited: Bhutan).

### Policy implications and suggested directions
- Domestic policies matter: untargeted subsidies that keep electricity prices artificially low can encourage excessive mining activity, exacerbating financial distortions and environmental costs.
- Macro and regulatory interactions:
  - Capital controls and weak domestic investment opportunities may make mining an outlet for capital flight; policymakers should consider how capital-account measures interact with crypto mining incentives.
- Future research avenues suggested by authors:
  - Use the framework to examine regulatory changes, mining bans, and the link between mining intensity and illicit financial flows.

*Source: What Drives Crypto Mining? Evidence from Hardware Imports — Working Paper No. WP/2026/### (selected content: 2.2, Table 1, Table 3, Table 4, Appendices A.1–A.6; sample period and figures as provided).*

---


_Source: https://www.imf.org/-/media/files/publications/wp/2026/english/wpiea2026146.pdf_