## Decrypting Crypto: How to Estimate International Stablecoin Flows — Working Paper No. WP/2025/141

## Source details

**Canonical URL:** [Decrypting Crypto: How to Estimate International Stablecoin Flows — Working Paper No. WP/2025/141](https://www.imf.org/-/media/files/publications/wp/2025/english/wpiea2025141-source-pdf.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/wp/2025/english/wpiea2025141-source-pdf.pdf.md)
- [Structured JSON version](/-/media/files/publications/wp/2025/english/wpiea2025141-source-pdf.pdf.json)

---

### Descriptive Statistics (Blockchains, Domains, Wallet Geography)
- Blockchain sample (as reported):
  - Ethereum: Genesis block 2015, Jul 30; Latest Block 21,525,890; #Transactions 2,639,611,278.
  - Binance Smart Chain: Genesis block 2020, Aug 29; Latest Block 45,369,482; #Transactions 6,523,262,103.
  - Optimism: Genesis block 2021, Jan 14; Latest Block 130,045,411; #Transactions 432,212,106.
  - Arbitrum: Genesis block 2021, May 28; Latest Block 290,687,173; #Transactions 1,222,934,534.
  - Base: Genesis block 2023, Jun 15; Latest Block 24,450,126; #Transactions 1,419,068,646.
  - Linea: Genesis block 2023, Jul 06; Latest Block 14,022,234; #Transactions 240,821,235.
  - Total #Transactions across chains: 12,477,909,902.
- Observations:
  - Ethereum longest-running (genesis 2015, Jul 30) but produces blocks more slowly (example: Ethereum ~every 12 seconds; Arbitrum ~every 0.25 seconds), so produced fewer blocks than some more recent blockchains.
  - Transaction counts per chain range from hundreds of millions to several billion; total exceeds 12 billion.
  - Dune Analytics identified as a commercial blockchain data provider.
- Domain name data:
  - Total domain names obtained: 5,933,958.
  - Breakdown by Name Service:
    - Ethereum Name Service: 3,413,426 registered domains.
    - Uxlink: 875,564 registered domains.
    - Spaceid: 474,021 registered domains.
    - Basenames: 437,329 registered domains.
    - Linea ENS Subdomains: 429,956 registered domains.
    - Arbid: 303,632 registered domains.
  - Observation: majority of the nearly 6 million domain names are ENS domains.

### Methodology for Estimating Geographic Region of Self‑Custodial Wallets (Overview)
- Regional partition used for classification:
  - Africa and the Middle East
  - Asia and the Pacific
  - Europe
  - North America
  - Latin America and the Caribbean
- Two sources of labeled regional information for wallets:
  1. Domain names assigned to wallets (e.g., ENS) — use an LLM to infer country, language, region and provide short reason; restrict region output to {Africa and the Middle East, Asia and Pacific, Europe, Latin America and Caribbean, North America, unclassified}.
  2. Frequent transactions with centralized exchanges targeting specific regional markets — classify wallet as belonging to a region if more than 90% of its transactions with centralized exchanges are with exchanges of that region.
- Labeled subset used to train a machine learning classifier: Gradient Boosted Decision Tree with Yggdrasil Decision Forests (YDF) in Python.

### Generating Training Data from Domain Names (LLM approach)
- LLM advantages and limitations:
  - Advantages: scales to millions of domains; ENS supports Unicode Technical Standard 51 allowing Arabic, Chinese, Korean, Japanese scripts and emojis, providing linguistic markers.
  - Limitations: users may choose names associated with other cultures/regions; migrant populations may select country-of-origin names; LLMs can misinterpret or exhibit biases (example: German names defaulted to Germany rather than Austria or Switzerland).
- LLM response settings preserved from source:
  - Maximum number of tokens: 150.
  - temperature: 0.4.
  - top_p: 0.4.
- Handling wallets with multiple domains:
  - Wallet region assigned as majority region among LLM’s guesses for its domains, excluding “unclassified” responses.
  - Example logic preserved: wallet with 10 domains, 6 unclassified, 3 Europe, 1 North America → assigned to Europe.

### Generating Training Data from Regional Exchanges
- Rationale: regionally focused CEXs (example: “Indodax” markets itself as an “Indonesian Bitcoin and Crypto Exchange” with a website entirely in Indonesian) imply wallet regionality.
- Classification rule: wallet classified as belonging to a region if >90% of its transactions with CEXs are with exchanges of that region.
- Reference to comprehensive list of regionally classified exchanges in Table 12 (appendix; table contents not reproduced here).

### Features and Model Training Procedure
- Training sample restriction: include only wallets that have initiated at least 50 transactions.
- Features used:
  - Time-of-day features: percentage of transactions in each hour; condensed by fitting a third degree polynomial and using coefficients as features (third degree chosen to accommodate both a minimum and a maximum).
  - Daylight Savings Time (DST) features: activity profiles during DST months vs remainder of year to exploit regional DST heterogeneity.
  - Top 5 centralized exchanges: counts of transactions with exchanges; categorical variables for most used exchange, second most used, etc.
  - Top 10 ERC-20 tokens: counts per wallet; categorical variables for top 10 tokens.
  - Smart contract namespaces: group contracts by namespace (Dune Analytics grouping); counts per wallet; categorical variables for top 10 namespaces.
- Training and evaluation:
  - Data split: training set 90%, testing set 10%.
  - Train 3 models combined to estimate region in {Africa and Middle East, Asia and Pacific, Europe, Latin America and Caribbean, North America}:
    - Model 1: assigns one of three grouped classifications: {North America, Latin America and Carribbean} OR {Africa and Middle East, Europe} OR Asia and Pacific (maximizes usefulness of time-of-day features).
    - Model 2: splits {North America, Latin America and Carribbean}.
    - Model 3: splits {Africa and Middle East, Europe}.
  - Validation: YDF automatically splits off validation data to avoid overfitting.
  - Class-imbalance handling: balanced sample weights—each observation weighted inversely proportional to class frequency.

### Classification of Arbitrary Self‑Custodial Wallets (Application)
- For out-of-sample wallets:
  - Compute same features as training set.
  - Restrict to wallets with at least 50 transactions.
  - Apply trained model to predict most likely region.
- Key identifying assumption:
  - Conditional on training features, wallets of a particular region in training follow same data-generating process as out-of-sample wallets (no systematic difference in hours of typical activity, etc.).

### Note on Results
- Source excerpt ends at the heading "5 Results"; detailed results from that section not included in provided excerpt.

---

### Training Data: Domain‑Name and Regional‑Exchange Labeled Wallets (Section 5.1)
- Domain-name LLM classification outputs: “unclassified”, “Africa and Middle East”, “Asia and Pacific”, “Europe”, “Latin America and Caribbean”, “North America”.
- Time‑zone offsets used to validate domain-name classifications via activity-profile aggregation:
  - Africa and the Middle East: UTC+2 (Central African time zone)
  - Asia and Pacific: UTC+8 (centered on China, parts of Indonesia, Malaysia, Singapore and Western Australia)
  - Europe: UTC+2 (Central European Summer Time)
  - Latin America and Caribbean: UTC-4 (parts of Brazil, Chile and Venezuela)
  - North America: UTC-6 (Central Time in the US)
- Example LLM classifications (selected):
  - vucoworld — Region: unclassified; Country: unclassified; Language: unclassified.
  - الكيميويات — Region: Africa and Middle East; Country: unclassified; Language: Arabic.
  - pijiu — Region: Asia and Pacific; Country: China; Language: Chinese.
  - philippzinner — Region: Europe; Country: Germany; Language: German.
  - laplazart — Region: Latin America and Caribbean; Country: Mexico; Language: Spanish.
  - lakings — Region: North America; Country: United States; Language: English.
- Validation finding: aggregate activity profiles show nighttime dip and peak roughly 10 AM to 10 PM local time across regions, supporting reasonable regional assignments from LLM domain classifications.

### Regional‑Exchange Labeled Wallets and DST Validation (Section 5.1.2)
- Method: identify wallets interacting with regionally focused CEXs and validate via activity profiles.
- DST findings:
  - North America: activity profile offsets by roughly one hour between March–October and November–February, consistent with one-hour DST shift.
  - Asia and Pacific: no significant DST-related difference (consistent with most countries in region not observing DST).
- Interpretation: DST detection supports correctness of regional wallet classifications.

### Training‑Data Coverage (Table 4 summary)
- Total wallets with identified region: 346,201
- Identification split: three quarters via domain names and one quarter via regional CEX usage.
- Domain-name identifications: 260,655 wallets out of almost 6 million (roughly 4.4%).
- Regional breakdown (Domain; Regional CEX; Total; % of Total):
  - Africa and Middle East — Domain: 19,159; Regional CEX: 10,466; Total: 29,625; % of Total: 8.6%
  - Asia and Pacific — Domain: 124,727; Regional CEX: 49,988; Total: 174,715; % of Total: 50.5%
  - Europe — Domain: 90,509; Regional CEX: 2,291; Total: 92,800; % of Total: 26.8%
  - Latin America and Caribbean — Domain: 3,723; Regional CEX: 2,090; Total: 5,813; % of Total: 1.7%
  - North America — Domain: 22,537; Regional CEX: 20,711; Total: 43,248; % of Total: 12.5%
  - Total — Domain: 260,655; Regional CEX: 85,546; Total: 346,201; % of Total: 100.0%
- Note: balanced sample weights employed to account for class imbalance.

### Classification Model Performance (Section 5.1.3)
- Model: Gradient boosted decision tree classifier.
- Confusion-matrix highlights (testing data):
  - Asia and Pacific predicted with 69.8% accuracy.
  - Africa and the Middle East predicted with 45.8% accuracy (worst-performing).
  - Random uniform guess baseline would be 20% per cell.
- Interpretation:
  - Predictive performance indicates distinctive on-chain behavioral patterns by region and supports LLM domain-name classifications.
  - Error asymmetries: tendency to misclassify wallets as Asia and Pacific and Europe due to majority-class effects (mitigated by balanced weights); some misclassifications may correct initial LLM biases (e.g., African/Middle East wallets misclassified as North America at 11.8% vs 3.6% in reverse).

---

### Estimating Region of Arbitrary Self‑Custodial Wallets (Section 5.2)
- Approach:
  - Use predicted probability distributions across regions rather than single most-likely label to improve accuracy and stability.
- Coverage and regional estimates (Table 5):
  - Total self-custodial wallets transferring stablecoins in sample: 20,174,942
  - Predicted regional wallet counts and percentages:
    - Africa and Middle East — # of Wallets: 3,010,170; % of Total: 13.59
    - Asia and Pacific — # of Wallets: 6,945,933; % of Total: 31.36
    - Europe — # of Wallets: 3,231,460; % of Total: 14.59
    - Latin America and Caribbean — # of Wallets: 2,556,156; % of Total: 11.54
    - North America — # of Wallets: 4,431,223; % of Total: 20.00
    - Total — # of Wallets: 20,174,942; % of Total: 100.00
  - Manual assignment: handful of wallets belonging to notable entities manually assigned (appendix E reference).
- Key 2024 summary statistics (leading into Section 6):
  - Total stablecoin transactions mapped: around 138 million
  - Total stablecoin volume: $2,019 billion
  - Average transaction size: $14,630
  - Total number of wallets involved: 14.6 million
    - Centralized exchange wallets: 10.4 million
    - Self-custodial wallets: 4.2 million
  - Volume split by counterparty type:
    - Flows entirely between self-custodial wallets: $309bn
    - Flows between self-custodial wallets and centralized exchanges: $1,141bn
    - Flows between centralized exchanges: $569bn
  - Self-custodial wallets present on at least one side of transactions comprising 72% of stablecoin transaction volume.

### Stablecoin and Regional Heterogeneity (Table 6)
- By region, USDC vs USDT volumes and shares:
  - Africa and Middle East — USDC: $8 bn (42.7%); USDT: $11.5 bn (57.3%)
  - Asia and Pacific — USDC: $179 bn (42.0%); USDT: $247 bn (58.0%)
  - Europe — USDC: $167 bn (50.0%); USDT: $167 bn (50.0%)
  - Latin America and Caribbean — USDC: $68 bn (43.3%); USDT: $88.5 bn (56.7%)
  - North America — USDC: $273 bn (61.3%); USDT: $172 bn (38.7%)
- Interpretation: USDT more popular in regions with more emerging markets; USDC more popular in regions with more advanced economies.

### Transaction Size Heterogeneity by Region (Table 7)
- Average and median transaction sizes by region:
  - Africa and Middle East — Average: $13,108; Median: $100
  - Asia and Pacific — Average: $11,493; Median: $94
  - Europe — Average: $18,878; Median: $200
  - Latin America and Caribbean — Average: $14,005; Median: $51
  - North America — Average: $35,016; Median: $101
- Note: averages significantly exceed medians, indicating fat-tailed distributions.

### Exchange Interaction Preferences by Region (Table 8)
- Percentage of volume with Coinbase vs Binance by region:
  - Africa and Middle East — Coinbase: 25.7; Binance: 74.3
  - Asia and Pacific — Coinbase: 16.9; Binance: 83.1
  - Europe — Coinbase: 33.7; Binance: 66.3
  - Latin America and Caribbean — Coinbase: 27.7; Binance: 72.3
  - North America — Coinbase: 54.0; Binance: 46.0
- Interpretation: emerging-market regions favor Binance; advanced-economy regions favor Coinbase.
- Notes and exclusions:
  - Exclude stablecoin transactions where sender and receiver are the same CEX (operational flows).
  - Exclude stablecoin flows with values less than 1 cent to avoid “dusting attacks”.

### Preview of 2024 Regional Flow Patterns (Section 6 introduction)
- Gross regional stablecoin flows (self-custodial wallets) — largest to smallest:
  - Asia and Pacific: $156.28 bn
  - North America: $118.26 bn
  - Latin America and Caribbean: $66.39 bn (smallest)
- Relative to GDP (using WEO 2023 regional GDP aggregations), gross flows/GDP:
  - Africa and Middle East: 1.5%
  - Latin America and Caribbean: 1.4%
  - Asia and Pacific: 0.4%
  - Europe: 0.4%
  - North America: 0.4%
- Net flows: virtually all stablecoin net flows are outflows from North American wallets ($21.54 bn) to other regions.
- Role of CEXs:
  - Binance processes more volume than any single geographic region of self-custodial wallets.
  - Binance and Other CEXs facilitate majority of inflows of stablecoins into self-custodial wallets (on‑ramping).
  - Coinbase receives far more inflows from self-custodial wallets than it sends to them (off‑ramping), reflecting differing geographic client bases and functional roles.

---

### Dividing Centralized Exchange (CEX) Flows into Regional Flows (Section 6.3)
- Allocation method:
  - For each CEX, calculate percentage of its total stablecoin volume transacted with self-custodial wallets in each region.
  - Use these percentages to allocate all flows involving the exchange across regions.
- Key assumption:
  - Geographic distribution of self-custodial wallets interacting with the exchange approximates geographic distribution of the exchange’s users.
- Advantage vs web-traffic approaches:
  - Does not assume CEX users do not use VPNs.
  - Accounts for regional differences in transaction sizes by using volume data.
- Outcome:
  - Allocation produces regional gross and net flow breakdowns similar to flows between self-custodial wallets, indicating interactions of self-custodial wallets with CEXs resemble interactions with other self-custodial wallets.
- Regional patterns and key statistics:
  - Largest regions in absolute stablecoin flows: North America and Asia and Pacific.
  - Regions with larger flows relative to GDP: Africa and Middle East (6.7%) and Latin America and Caribbean (7.7%).
  - Inter-region flows more important than within-region flows → stablecoins used for international capital flows and remittances.
  - Net flows: most significant net flow is outflow from North America totalling $54.06bn to all other regions.
- Mechanism reinforcing North America as net exporter:
  - Eligible users transfer $1 in fiat to issuers who mint equivalent stablecoins; when market price > $1, North American participants can exchange fiat for newly issued stablecoins and sell on secondary markets, reinforcing net outflow (see Makarov and Schoar (2022)).

---

### Country Focus: China (Section 6.4)
- Motivation:
  - Anecdotal reports of high crypto activity despite official bans; Chainalysis assumptions (no VPN use) likely violated for China.
- Classification:
  - Binance most important CEX for China in stablecoin flows, despite Binance.com requiring VPN access from China.
  - Model separating China vs Asia and Pacific (ex. China) achieves 79% accuracy.
- Flows involving China (2024):
  - Gross inflows: $84.03bn
  - Gross outflows: $69.05bn
  - Within‑country flows: $4.77bn
  - Interpretation: within-country flows small relative to in- and outflows → flows mainly facilitate international capital flows rather than domestic payments.
- Key counterparties and bilateral/net flows:
  - Binance facilitates $32.27bn into Chinese self-custodial wallets and $21.00bn from Chinese self-custodial wallets.
  - Other CEXs also significant; Coinbase gross flows involving China relatively small.
  - Regional counterpart gross flows for China (Asia and Pacific ex. China): inflows $6.87bn; outflows $6.74bn (relatively balanced).
  - Net bilateral flows into Chinese self-custodial wallets total $18.58bn comprised of:
    - Binance: $11.28bn
    - Other CEXs: $5.78bn
    - North America: $1.52bn
  - Major outflows from Chinese self-custodial wallets mostly flow to Coinbase: $3.02bn.
  - Net inflows into China equal 4.4% of the current account surplus (current account surplus used: $424bn in 2024).

---

### Economic Drivers of Stablecoin Flows — Exchange Rates and Events (Section 7)
- Hypothesis and regression setup (Section 7.1):
  - Net flows from North America to other regions tracked daily from January 1, 2022 to December 31, 2024.
  - Estimated regression: Net Flows vs NA_{r,t} = β1 VIX_t + β2 Broad Dollar_t + β3 Crypto F&G_t + α_{r,Q} + Weekend + ε_{r,t}
    - VIX: market volatility index.
    - Broad Dollar: trade-weighted index of exchange rates vs US-Dollar.
    - Crypto F&G: Crypto Fear & Greed Index.
    - Region-by-quarter fixed effects and weekend dummy included; weekend VIX and Broad Dollar interpolated linearly to align with 24/7 activity.
- Empirical results (Table 9 summary):
  - Global / China reported entries preserved as in source:
    - VIX -0.018 0.098 (0.068) (0.053)
    - Broad Dollar 0.181 ∗∗ (0.085)
    - USD/CNY 0.284 ∗∗∗ (0.095)
    - Crypto F&G 0.025 -0.040 (0.042) (0.045)
    - Region×Quarter FE ✓
    - Quarter FE ✓
    - Weekend FE ✓✓
    - Observations 4·1096 1096
    - R-squared 0.3100.200
    - F-statistic 40.58016.813
  - Interpretation:
    - Broad Dollar coefficient 0.181 ∗∗ (0.085): stronger U.S. dollar associated with significant increase in outflows of stablecoins from North America to other regions.
    - For China, USD/CNY coefficient 0.284 ∗∗∗ (0.095): stronger dollar versus Renminbi associated with increased stablecoin flows into China.
    - VIX and Crypto F&G not significant in global regressions.
  - Note: residuals tested for stationarity at the 0.01 level; additional specification with lag in Appendix F.

### Impact of March 2023 US Banking Crisis (Section 7.2)
- Rationale: stablecoin issuance/redemption depends on banks managing fiat reserves; failure of crypto-related banks disrupted fiat-side operations.
- Event-study / difference-in-differences:
  - Compare weekly flows from North America (treated) vs other regions (controls) for five weeks pre/post crisis.
  - Treatment day: March 10th 2023 (Silicon Valley Bank collapse and FDIC receivership).
  - Regression: Flows from Region_{r,t} = α_r + α_t + sum_{τ=-5, τ≠-1}^{τ=5} β_τ · Treated_r · 1{τ=t} + ε_{r,t}
- Results:
  - Parallel pre-trends observed.
  - On impact, stablecoin flows originating from North America drop dramatically: decrease of almost 10 standard deviations relative to controls.
  - Flows normalize in subsequent weeks, returning to normal.
- Interpretation: March 2023 banking-sector failures caused a substantial but temporary disruption to stablecoin flows from North America.

---

### Comparison with Chainalysis Dataset (Section 8)
- Methodological differences:
  - Chainalysis: uses web-traffic (Similarweb) to infer CEX user geography; assumes no VPN use and uniform average transaction sizes across countries.
  - This paper: does not rely on web-traffic or VPN non-use; uses volume and wallet interaction data; covers subset of blockchains but likely more exhaustive within those chains.
  - Chainalysis “indirect” category likely captures only a fraction of self-custodial transactions.
- Aggregate quantity comparison:
  - This paper’s dataset: $2 trillion
  - Chainalysis: $1.7 trillion
- Agreement:
  - Both datasets: USDT more prevalent in EM regions; USDC favored in AE regions.
  - Asia and Pacific exhibits largest stablecoin flows; Africa and the Middle East and Latin America and the Caribbean smaller absolute volumes.
  - Direct (CEX-to-CEX) estimates broadly agree: net outflows originate from North America to other regions.
- Direct vs Indirect category comparison:
  - Direct: no significant disagreement between datasets.
  - Indirect: severe disagreement.
    - This paper’s indirect category: stablecoins primarily flow from North America to other regions.
    - Chainalysis indirect category: suggests stablecoins flow out of all regions into North America.
    - Discrepancy puzzling given similar intended capture of transfers involving self-custodial wallets and CEXs; raises questions about Chainalysis indirect category reliability.
- China-specific disagreement:
  - VPN use in China likely causes Chainalysis undercount.
  - Comparative figures:
    - Gross flows involving China: This paper $153 billion vs Chainalysis $28 billion (ratio 5.5).
    - Net flows involving China: This paper $18 billion vs Chainalysis $0.18 billion (ratio 100).

---

### Correlations Between Datasets (Section 8.2)
- Gross-flow correlations (weekday series Jan 1 2024–Dec 31 2024):
  - Very large positive correlations between outflows in our dataset and Chainalysis, ranging from 0.78 to 0.96.
  - Regional outflow correlations (Our dataset vs Chainalysis):
    - Africa and Middle East: Direct 0.88, Indirect 0.79, Direct + Indirect 0.91
    - Asia and Pacific: Direct 0.89, Indirect 0.85, Direct + Indirect 0.94
    - Europe: Direct 0.90, Indirect 0.89, Direct + Indirect 0.96
    - Latin America and Caribbean: Direct 0.86, Indirect 0.78, Direct + Indirect 0.87
    - North America: Direct 0.91, Indirect 0.88, Direct + Indirect 0.95
- Net-flow correlations and divergences:
  - High positive correlation for net flows in direct category between datasets.
  - Chainalysis indirect category correlations with this paper range from large negative to negligible — indicating disagreement in direction and dynamics of net flows.
  - Net-flow correlation table (Our dataset vs Chainalysis; format Direct, Indirect, Direct + Indirect, Chainalysis Direct vs Indirect):
    - Africa and Middle East: 0.27, -0.18, 0.10, -0.42
    - Asia and Pacific: 0.75, 0.01, 0.50, -0.13
    - Europe: 0.22, -0.15, 0.35, -0.27
    - Latin America and Caribbean: 0.60, -0.38, 0.09, -0.47
    - North America: 0.80, -0.22, 0.53, -0.39
  - Implication: Chainalysis indirect category disagrees with this paper (and Chainalysis direct) on direction and time-series dynamics of net flows.
- Inflow-outflow correlations within datasets:
  - Prior research: Chainalysis inflow-outflow correlations up to 99%.
  - This paper finds similarly large inflow-outflow correlations to Chainalysis for aggregate samples.
  - When computing inflow-outflow correlations purely between self-custodial wallets (possible in this paper but not Chainalysis), correlations are significantly lower than full-sample correlations, though still high — suggests exchange-related activity drives much of the high inflow-outflow symmetry.
  - Inflow correlations (Appendix Table 16) — Correlation of Inflows between Respective Categories in Our Dataset and Chainalysis:
    - Africa and Middle East: Direct 0.89, Indirect 0.83, Direct + Indirect 0.93
    - Asia and Pacific: Direct 0.93, Indirect 0.87, Direct + Indirect 0.95
    - Europe: Direct 0.92, Indirect 0.90, Direct + Indirect 0.96
    - Latin America and Caribbean: Direct 0.89, Indirect 0.84, Direct + Indirect 0.91
    - North America: Direct 0.81, Indirect 0.83, Direct + Indirect 0.94
  - Inflow-outflow correlations by category (Appendix Table 17) — Our Dataset vs Chainalysis; entries listed as: Our Dataset Direct, Indirect, Total, Self-custodial; Chainalysis Direct, Indirect, Total:
    - Africa and Middle East: 0.99, 0.98, 0.99, 0.89; 0.96, 0.96, 0.98
    - Asia and Pacific: 0.94, 0.98, 0.98, 0.88; 0.97, 0.98, 0.99
    - Europe: 0.99, 0.98, 0.99, 0.84; 0.99, 0.99, 1.00
    - Latin America and Caribbean: 0.99, 0.99, 0.99, 0.91; 0.97, 0.97, 0.98
    - North America: 0.94, 0.93, 0.95, 0.66; 0.94, 0.93, 0.96
  - Note: within-region flows excluded to avoid mechanical inflow-outflow correlation.
- Implication for measurement and interpretation:
  - High gross-flow correlations show strong agreement on magnitudes of outflows.
  - Discrepancies in net-flow correlations—especially Chainalysis indirect—highlight methodological/classification differences that affect inferred directionality and dynamics.
  - Difference in inflow-outflow correlations between CEX-involving samples and self-custodial-only samples implies exchange-related wallet activity drives much of aggregate inflow-outflow symmetry.

*Source: wpiea2025141-source-pdf - 3.1 Descriptive Statistics; 5.1 Training Data; 6.3 Dividing Centralized Exchange Flows into Regional Flows; 6.4 Country Focus: China; 7 Economic Drivers of Stablecoin Flows; 8 Comparison with Chainalysis; 8.2 Correlations*

### 3.1 Descriptive Statistics

### 3.1 Descriptive Statistics

### Description of blockchain data
- Ethereum: Genesis block 2015, Jul 30; Latest Block 21,525,890; #Transactions 2,639,611,278.
- Binance Smart Chain: Genesis block 2020, Aug 29; Latest Block 45,369,482; #Transactions 6,523,262,103.
- Optimism: Genesis block 2021, Jan 14; Latest Block 130,045,411; #Transactions 432,212,106.
- Arbitrum: Genesis block 2021, May 28; Latest Block 290,687,173; #Transactions 1,222,934,534.
- Base: Genesis block 2023, Jun 15; Latest Block 24,450,126; #Transactions 1,419,068,646.
- Linea: Genesis block 2023, Jul 06; Latest Block 14,022,234; #Transactions 240,821,235.
- Total #Transactions across chains: 12,477,909,902.
- Observations from the source text:
  - Ethereum is the longest running blockchain in the sample (genesis 2015, Jul 30) but produces blocks more slowly (example: Ethereum produces a new block approximately every 12 seconds, while Arbitrum produces a block every 0.25 seconds), so it has produced fewer blocks than some more recent blockchains such as Optimism or Arbitrum.
  - The number of transactions for each blockchain ranges from hundreds of millions to several billion, with the total number exceeding 12 billion transactions.
- Note: Dune Analytics is identified as a commercial blockchain data provider in the text.

### Description of domain name data
- Total domain names obtained: 5,933,958.
- Breakdown by Name Service:
  - Ethereum Name Service: 3,413,426 registered domains.
  - Uxlink: 875,564 registered domains.
  - Spaceid: 474,021 registered domains.
  - Basenames: 437,329 registered domains.
  - Linea ENS Subdomains: 429,956 registered domains.
  - Arbid: 303,632 registered domains.
- Observation: The majority of the nearly 6 million domain names are ENS domains.

### Methodology for estimating the geographic region of self-custodial wallets
- World divided into five regions for classification:
  - Africa and the Middle East
  - Asia and the Pacific
  - Europe
  - North America
  - Latin America and the Caribbean
- Two approaches for obtaining geographic information for a subset of wallets:
  1. Domain names assigned to wallets (e.g., ENS): use a large language model (LLM) to infer linguistic and cultural markers—such as language, script, or regional references—that suggest a wallet’s likely region.
  2. Frequent transactions with centralized exchanges targeting specific regional markets: assume a wallet predominantly interacting with a region-focused exchange is likely from that region.
- These two methods produce an ad hoc regional classification for a subset of wallets, which is then used as labeled training data for a machine learning classification model.
- Machine learning model used: Gradient Boosted Decision Tree implemented with the Yggdrasil Decision Forests (YDF) library in Python.

### Generating training data
- By analyzing domain names:
  - LLM advantages:
    - Scales to analyze millions of domains quickly and efficiently.
    - ENS supports Unicode Technical Standard 51, allowing diverse characters (including Arabic, Chinese, Korean, Japanese scripts, and emojis), providing clear linguistic markers in many cases.
    - In some instances, inferring region can be more straightforward than pinpointing a specific country.
  - LLM limitations and caveats:
    - Users may reside in one region but choose names culturally or linguistically associated with another (e.g., English names or U.S. cultural references used globally).
    - Migrant populations may select domain names that evoke country of origin.
    - LLMs can misinterpret inputs or exhibit biases (example: German domain names may be defaulted to Germany rather than Austria or Switzerland).
  - LLM prompt (essence preserved from source):
    - Instructs the LLM to classify ENS domain names by country, language, region, and provide a short reason; to consider references to culture, language, localities, memes; to be mindful of language commonly used in crypto and web3; to output ‘unclassified’ if unable to classify a particular attribute; to restrict region output to the set {Africa and the Middle East, Asia and Pacific, Europe, Latin America and Caribbean, North America, unclassified}.
  - LLM response settings:
    - Maximum number of tokens: 150 (roughly 110-120 words in English).
    - temperature: 0.4.
    - top_p: 0.4.
  - Handling wallets with multiple domains:
    - Assign the wallet’s region as the majority region among the LLM’s guesses for its domains, excluding “unclassified” responses.
    - Example logic: wallet owns 10 domains, 6 unclassified, 3 Europe, 1 North America → assigned to Europe.
- By linking wallets to regional exchanges:
  - Rationale:
    - Some centralized exchanges focus on particular regional markets (example given: “Indodax” markets itself as an “Indonesian Bitcoin and Crypto Exchange” with a website entirely in Indonesian).
    - A self-custodial wallet that frequently receives or sends money to a region-focused exchange is likely from that region.
  - Classification rule:
    - A self-custodial wallet is classified as belonging to a particular region if more than 90% of its transactions with centralized exchanges are with centralized exchanges of that particular region.
  - The text references a comprehensive list of regionally classified exchanges in Table 12 in the appendix (table contents not reproduced here).

### Training the classification model
- Sample restriction for training:
  - Include only wallets that have initiated at least 50 transactions to ensure enough observations per wallet.
- Features used to train the classification model:
  - Time of Day Features:
    - For every wallet, calculate the percentage of transactions conducted within every hour of the day.
    - To condense data and avoid overfitting, estimate a third degree polynomial describing the activity profile and use the coefficients as features (third degree polynomial chosen to be able to accompany both a minimum and a maximum in the activity profile).
  - Daylight Savings Time Features:
    - For every wallet, calculate activity profile during the months of daylight savings time and during the remainder of the year to exploit regional heterogeneity in the use of daylight savings time (DST common in Europe and North America; not common in many other regions and non-existent in Asia).
  - Top 5 Centralized Exchanges:
    - For every wallet, count the number of transactions with centralized exchanges (withdrawals and deposits).
    - Sort descending by exchange name and create categorical variables for the most used centralized exchange, second most used, etc.
  - Top 10 ERC-20 Tokens and Smart Contract Namespaces:
    - For top 10 ERC-20 tokens: count per wallet the number of transactions using ERC-20 contracts, sort descending, create categorical variables for the top 10 most used.
    - For smart contract namespaces: group smart contracts by “namespace” (groups smart contracts that offer similar functionalities as provided by Dune Analytics), count per wallet, sort descending, create categorical variables for the top 10 most interacted namespaces.
    - Rationale: ERC-20 level offers granularity; namespace groups exploit heterogeneity in functionality.
- Training and evaluation procedure:
  - Data split: training data set 90% of the data; testing data set 10% of the data.
  - Train 3 models that are combined to estimate the likely region from the set {Africa and Middle East, Asia and Pacific, Europe, Latin America and Caribbean, North America}.
  - Decision-tree structure and modeling approach (as described in the source):
    - First model assigns one of three grouped classifications: {North America, Latin America and Carribbean} OR {Africa and Middle East, Europe} OR Asia and Pacific.
      - This grouping maximizes usefulness of time-of-day based features because grouped regions largely share the same time zones.
    - Second model splits {North America, Latin America and Carribbean} into its respective regions.
    - Third model splits {Africa and Middle East, Europe} into its respective regions.
    - The further models exploit other features to distinguish between respective regions.
  - Validation and class imbalance handling:
    - For each of the three models, the YDF package automatically splits off some of the data for validation to avoid overfitting.
    - Address imbalance in training data by calculating balanced sample weights—each observation weighted inversely proportional to its class frequency—ensuring underrepresented classes contribute equally (references cited in source: King and Zeng(2001), He and Garcia (2009)).
- Classification of arbitrary self-custodial wallets:
  - For out-of-sample wallets, calculate the same features used in training and restrict to wallets with at least 50 transactions (to align with training restriction).
  - Apply the trained model to predict the most likely region a wallet belongs to.
  - Key identifying assumption:
    - Conditional on the features used to train the model, wallets of a particular region in the subset used for training follow the same data generating process as out-of-sample wallets the model is applied to (example: assume no systematic difference between a North American wallet in the training subset and an out-of-sample North American wallet with respect to hours of typical activity).

### Results
- The source content ends at the heading "5 Results"; results content is not included in the provided excerpt.

*Source: wpiea2025141-source-pdf - 3.1 Descriptive Statistics*

### 5.1 Training Data

### 5.1 Training Data

### 5.1.1 Domain Names
- Method:
  - LLM classifies domain names into outputs: “unclassified”, “Africa and Middle East”, “Asia and Pacific”, “Europe”, “Latin America and Caribbean”, and “North America”.
  - Validation: for each wallet with a classified domain name, collect timestamps of every transaction, compute wallet-specific activity profile by hour of day (normalized % of transactions per hour), offset UTC timestamps to a regional proxy time zone, and aggregate by region.
  - Time zone offsets used:
    - Africa and the Middle East: UTC+2 (Central African time zone)
    - Asia and Pacific: UTC+8 (centered on China, parts of Indonesia, Malaysia, Singapore and Western Australia)
    - Europe: UTC+2 (Central European Summer Time)
    - Latin America and Caribbean: UTC-4 (centering on parts of Brazil, Chile and Venezuela)
    - North America: UTC-6 (Central Time in the US)
- Example classifications (Table 3), with provided reasons:
  - vucoworld — Region: unclassified; Country: unclassified; Language: unclassified. Reason: The term ‘vucoworld’ does not clearly reference a specific country, language, or region.
  - الكيميويات — Region: Africa and Middle East; Country: unclassified; Language: Arabic. Reason: Domain name is in Arabic and translates to ‘chemicals’.
  - pijiu — Region: Asia and Pacific; Country: China; Language: Chinese. Reason: ‘Pijiu’ means ‘beer’ in Chinese.
  - philippzinner — Region: Europe; Country: Germany; Language: German. Reason: Suggests a German personal name.
  - laplazart — Region: Latin America and Caribbean; Country: Mexico; Language: Spanish. Reason: Suggests connection to ‘La Plaza’.
  - lakings — Region: North America; Country: United States; Language: English. Reason: Likely refers to the Los Angeles Kings hockey team.
- Validation findings:
  - Aggregate activity profiles (Figure 3a) show a distinctive dip in activity during nighttime and highest activity roughly from 10 AM to 10 PM local time across regions, supporting that LLM domain-name classifications produce reasonable regional assignments in aggregate.

### 5.1.2 Regional Exchange Users
- Method:
  - Identify wallets that interact with regionally focused centralized exchanges (CEXs) and validate using the same activity profile approach as domain names.
- Validation findings:
  - Activity profiles for wallets identified through regional exchanges are depicted in Figure 3b.
  - Daylight Savings Time effect (Figure 4):
    - North America: activity profile offsets by roughly one hour between March–October (Standard Time window used in paper description) and November–February (Daylight Saving Time window), consistent with a one-hour DST shift.
    - Asia and Pacific: no significant difference in activity by time of day across the same windows, consistent with most countries in the region (except Australia and New Zealand) not observing DST.
  - Authors interpret DST detection as further evidence that wallet classifications by region are, on average, correct.
- Training-data coverage (Table 4):
  - Total wallets with identified region: 346,201
  - Identification split: three quarters via domain names and one quarter via usage of regional CEXs.
  - Domain-name identifications: 260,655 wallets out of almost 6 million (roughly 4.4%).
  - Regional breakdown:
    - Africa and Middle East — Domain: 19,159; Regional CEX: 10,466; Total: 29,625; % of Total: 8.6%
    - Asia and Pacific — Domain: 124,727; Regional CEX: 49,988; Total: 174,715; % of Total: 50.5%
    - Europe — Domain: 90,509; Regional CEX: 2,291; Total: 92,800; % of Total: 26.8%
    - Latin America and Caribbean — Domain: 3,723; Regional CEX: 2,090; Total: 5,813; % of Total: 1.7%
    - North America — Domain: 22,537; Regional CEX: 20,711; Total: 43,248; % of Total: 12.5%
    - Total — Domain: 260,655; Regional CEX: 85,546; Total: 346,201; % of Total: 100.0%
- Note: To account for class imbalance in training, the authors employ balanced sample weights (section 4.2 reference).

### 5.1.3 Classification Model
- Model:
  - Gradient boosted decision tree classifier trained on the labeled training data.
- Evaluation:
  - Confusion matrix on secluded testing data normalized by true-region rows shown in Figure 5.
  - Performance highlights:
    - Asia and Pacific predicted with 69.8% accuracy.
    - Africa and the Middle East predicted with 45.8% accuracy (worst-performing).
    - Even worst-performing predictions are considerably better than random (random uniform guess would show 20% per cell).
  - Interpretation:
    - Good predictive performance implies distinctive on-chain behavioral patterns by region and supports the LLM’s role in correctly classifying domain names for training data.
  - Error patterns (off-diagonal analysis):
    - Errors are not symmetric; there is a tendency for wallets to be more likely misclassified as Asia and Pacific and Europe—likely due to those regions being majority classes in training data (mitigated via balanced sample weights).
    - Some asymmetric “misclassifications” may correct initial LLM misclassification biases. Example: wallets truly from Africa and Middle East are misclassified as North America at 11.8%, whereas true North America wallets classified as Africa and Middle East are 3.6%—interpreted as reflecting immigrant populations and desirable correction by the model.
    - Misclassification symmetry is stronger for Asia and Pacific, possibly because majority-class bias is balanced by countervailing immigration flows.
  - Additional performance metrics and diagnostic graphs (ROC, AUC, etc.) provided in appendix D.

### 5.2 Estimating the Region of Arbitrary Self-Custodial Wallets
- Approach:
  - Use predicted probabilities for each region (full probability distribution) from the classification model rather than only the single most-likely predicted region to improve accuracy and stability.
- Coverage and regional estimates (Table 5):
  - Total self-custodial wallets transferring stablecoins in sample: 20,174,942
  - Predicted regional wallet counts and percentages:
    - Africa and Middle East — # of Wallets: 3,010,170; % of Total: 13.59
    - Asia and Pacific — # of Wallets: 6,945,933; % of Total: 31.36
    - Europe — # of Wallets: 3,231,460; % of Total: 14.59
    - Latin America and Caribbean — # of Wallets: 2,556,156; % of Total: 11.54
    - North America — # of Wallets: 4,431,223; % of Total: 20.00
    - Total — # of Wallets: 20,174,942; % of Total: 100.00
  - Manual assignment: a handful of wallets belonging to notable entities were manually assigned (appendix E reference).

- Key summary statistics for 2024 (overview leading into Section 6):
  - Total stablecoin transactions mapped: around 138 million
  - Total stablecoin volume: $2,019 billion
  - Average transaction size: $14,630
  - Total number of wallets involved: 14.6 million
    - Centralized exchange wallets: 10.4 million
    - Self-custodial wallets: 4.2 million
  - Volume split by counterparty type:
    - Flows entirely between self-custodial wallets: $309bn
    - Flows between self-custodial wallets and centralized exchanges: $1,141bn
    - Flows between centralized exchanges: $569bn
  - Self-custodial wallets are present on at least one side of transactions comprising 72% of stablecoin transaction volume.
- Stablecoin and regional heterogeneity (Table 6):
  - Africa and Middle East — USDC: $8 bn (42.7%); USDT: $11.5 bn (57.3%)
  - Asia and Pacific — USDC: $179 bn (42.0%); USDT: $247 bn (58.0%)
  - Europe — USDC: $167 bn (50.0%); USDT: $167 bn (50.0%)
  - Latin America and Caribbean — USDC: $68 bn (43.3%); USDT: $88.5 bn (56.7%)
  - North America — USDC: $273 bn (61.3%); USDT: $172 bn (38.7%)
  - Interpretation: USDT is more popular in regions with more emerging markets; USDC is more popular in regions with more advanced economies.
- Transaction size heterogeneity by region (Table 7):
  - Africa and Middle East — Average Transaction Size: $13,108; Median Transaction Size: $100
  - Asia and Pacific — Average Transaction Size: $11,493; Median Transaction Size: $94
  - Europe — Average Transaction Size: $18,878; Median Transaction Size: $200
  - Latin America and Caribbean — Average Transaction Size: $14,005; Median Transaction Size: $51
  - North America — Average Transaction Size: $35,016; Median Transaction Size: $101
  - Note: Averages significantly exceed medians, indicating fat-tailed distributions (extreme values).
- Exchange interaction preferences by region (Table 8):
  - Percentage of volume with Coinbase vs Binance:
    - Africa and Middle East — Coinbase: 25.7; Binance: 74.3
    - Asia and Pacific — Coinbase: 16.9; Binance: 83.1
    - Europe — Coinbase: 33.7; Binance: 66.3
    - Latin America and Caribbean — Coinbase: 27.7; Binance: 72.3
    - North America — Coinbase: 54.0; Binance: 46.0
  - Interpretation: Regions with more emerging markets favor Binance; regions with more advanced economies favor Coinbase.
- Notes and exclusions:
  - Stablecoin transactions where sender and receiver are the same centralized exchange are excluded (operational flows between deposit addresses, hot wallets, cold storage).
  - Stablecoin flows with values less than 1 cent are excluded to avoid skew from “dusting attacks”.
- Preview of 2024 regional flow patterns (Section 6 introduction):
  - Gross regional stablecoin flows (self-custodial wallets) — largest: Asia and Pacific ($156.28 bn); second: North America ($118.26 bn); smallest: Latin America and Caribbean ($66.39 bn).
  - Relative to GDP (percentages calculated as (inflows + outflows + within flows)/GDP using WEO 2023 regional GDP aggregations):
    - Africa and Middle East: 1.5%
    - Latin America and Caribbean: 1.4%
    - Asia and Pacific: 0.4%
    - Europe: 0.4%
    - North America: 0.4%
  - Net flows: virtually all stablecoin net flows are outflows from North American wallets ($21.54 bn) to other regions.
- Role of centralized exchanges in cross-border stablecoin flows (Section 6.2 preview and Figure 7):
  - Binance processes more volume than any single geographic region of self-custodial wallets.
  - Binance and Other CEXs facilitate the majority of inflows of stablecoins into self-custodial wallets across regions (on-ramping).
  - Coinbase receives far more inflows from self-custodial wallets than it sends to them (off-ramping), reflecting differing geographic client bases and potential functional roles in fiat–crypto flows.

*Source: 5.1 Training Data (content unit: wpiea2025141-source-pdf - 5.1 Training Data)*

### 6.3 Dividing Centralized Exchange Flows into Regional Flows

### 6.3 Dividing Centralized Exchange Flows into Regional Flows

### Method for allocating CEX flows to regions
- For each centralized exchange (CEX), calculate the percentage of its total stablecoin volume transacted with self-custodial wallets in each region.
- Use these percentages to allocate all flows involving the exchange across regions.
- Key assumption: the geographic distribution of self-custodial wallets interacting with the exchange approximates the geographic distribution of the exchange’s users.
- Advantage over web-traffic-based approaches (e.g., Chainalysis): does not assume CEX users do not use VPNs and inherently accounts for regional differences in transaction sizes by using volume data.

### Visual and comparative outcome
- The allocation produces regional gross and net stablecoin flow breakdowns (Figure 8).
- The regional breakdown for flows involving CEXs is very similar to flows between self-custodial wallets (Section 6), indicating interactions of self-custodial wallets with CEXs resemble interactions with other self-custodial wallets.

### Regional patterns and key statistics
- Largest regions in absolute stablecoin flows: North America and Asia and Pacific.
- Regions with larger flows relative to GDP: Africa and Middle East (6.7%) and Latin America and Caribbean (7.7%).
- Inter-region flows are more important than within-region flows, supporting the interpretation that stablecoins are used for international capital flows and remittances.
- For net flows, the most significant flow is an outflow from North America totalling $54.06bn to all other regions.
- Mechanism reinforcing North America as a net exporter:
  - Eligible users (likely more concentrated in North America due to greater fiat dollar access) transfer $1 in fiat to issuers, who mint equivalent stablecoins.
  - Arbitrage: when market price > $1, North American participants can exchange fiat for newly issued stablecoins and sell on secondary markets, reinforcing net outflow (see Makarov and Schoar (2022)).

---

### 6.4 Country Focus: China

### Motivation and classification
- China examined because anecdotal reports describe high crypto activity despite official bans, and Chainalysis assumptions (no VPN use) are likely violated for China.
- Binance appears to be the most important CEX for China in terms of stablecoin flows, despite Binance.com requiring VPN access from China.
- A model to separate wallets into China vs Asia and Pacific (ex. China) achieves 79% accuracy.

### Flows involving China (2024)
- Gross flows involving China:
  - Inflows: $84.03bn
  - Outflows: $69.05bn
  - Within-country flows: $4.77bn
- Interpretation: within-country flows are small relative to in- and outflows, implying stablecoin flows involving China mainly facilitate international capital flows rather than domestic payments.

### Key counterparties and bilateral/net flows
- Binance facilitates:
  - $32.27bn in flows into Chinese self-custodial wallets
  - $21.00bn in flows from Chinese self-custodial wallets
- “Other CEX” category also plays a significant role; gross flows involving Coinbase are relatively small.
- Regional counterpart flows for China (gross):
  - Asia and Pacific (ex. China) in- and outflows: $6.87bn and $6.74bn respectively (relatively balanced).
- Net bilateral flows into Chinese self-custodial wallets: $18.58bn total, comprised of:
  - Binance: $11.28bn
  - Other CEXs: $5.78bn
  - North America: $1.52bn
- Major outflows from Chinese self-custodial wallets mostly flow to Coinbase: $3.02bn.
- Net inflows of stablecoins into China amount to 4.4% of the current account surplus (using a current account surplus of $424bn in 2024).

---

### 7 Economic Drivers of Stablecoin Flows

### 7.1 Link between Stablecoins and Exchange Rates — methodology
- Hypothesis: net outflows from North America are linked to demand for U.S. dollars (e.g., when local currencies depreciate).
- Constructed panel time series of daily net stablecoin flows from North America to other regions from January 1, 2022, to December 31, 2024.
- Estimated regression:
  Net Flows vs NA_{r,t} = β1 VIX_t + β2 Broad Dollar_t + β3 Crypto F&G_t + α_{r,Q} + Weekend + ε_{r,t}
  - VIX: market volatility index.
  - Broad Dollar: trade-weighted index of exchange rates vs US-Dollar.
  - Crypto F&G: Crypto Fear & Greed Index (higher = bullish).
  - Region-by-quarter fixed effects and weekend dummy included.
  - Weekend VIX and Broad Dollar values interpolated linearly to align with 24/7 stablecoin activity.

### 7.1 Link — empirical results (Table 9)
- Global/regional regression highlights:
  - Broad Dollar coefficient: 0.181 ∗∗ (standard error (0.085)) — a stronger U.S. dollar is associated with a significant increase in outflows of stablecoins from North America to other regions.
  - VIX: -0.018 (0.068) and Crypto F&G: 0.025 (0.042) — no significant impact found for volatility or crypto sentiment on net stablecoin flows from North America.
- China-specific regression highlights:
  - USD/CNY coefficient: 0.284 ∗∗∗ (standard error (0.095)) — a stronger dollar versus the Renminbi is associated with increased stablecoin flows into China.
  - VIX (China column): 0.098 (0.053) and Crypto F&G: -0.040 (0.045) — not significant.
- Table 9 reported statistics (preserved exactly as in source):
  - Global / China column entries as presented:
    - VIX -0.018 0.098 (0.068) (0.053)
    - Broad Dollar 0.181 ∗∗ (0.085)
    - USD/CNY 0.284 ∗∗∗ (0.095)
    - Crypto F&G 0.025 -0.040 (0.042) (0.045)
    - Region×Quarter FE ✓
    - Quarter FE ✓
    - Weekend FE ✓✓
    - Observations 4·1096 1096
    - R-squared 0.3100.200
    - F-statistic 40.58016.813
  - Standard errors in parentheses. Significance: ∗ p<0.1, ∗∗ p<0.05, ∗∗∗ p<0.01.
  - Note: residuals tested for stationarity at the 0.01 level. Additional specification with lag in Appendix F.

### 7.2 Impact of the March 2023 US banking crisis
- Rationale: stablecoin issuance/redemption requires banks to manage fiat reserves and settle fiat legs; failure of crypto-related banks disrupted fiat-side operations.
- Event study / difference-in-differences:
  - Compare weekly flows originating from North America (treated) to flows from other regions (controls) over five weeks pre- and post-crisis.
  - Treatment day: March 10th 2023 (Silicon Valley Bank collapsed and was put under FDIC receivership).
  - Regression: Flows from Region_{r,t} = α_r + α_t + sum_{τ=-5, τ≠-1}^{τ=5} β_τ · Treated_r · 1{τ=t} + ε_{r,t}
- Results:
  - Parallel pre-trends observed, supporting validity of controls.
  - On impact, stablecoin flows originating from North America drop dramatically, exhibiting a decrease of almost 10 standard deviations relative to controls.
  - Flows normalize in subsequent weeks, returning to normal.
- Interpretation: banking sector failures in March 2023 caused a substantial but temporary disruption to stablecoin flows from North America.

---

### 8 Comparison with Chainalysis Dataset

### Methodological differences and coverage
- Chainalysis method: uses web-traffic data (Similarweb) to infer geographic locations of CEX users, focusing on on-chain transactions between CEXs.
  - Implicit assumptions: users do not use VPNs; average transaction sizes are uniform across countries.
- This paper’s dataset:
  - Does not rely on web-traffic assumptions or VPN non-use.
  - Uses volume and wallet interaction data to map flows between regions and between regions and specific CEXs.
  - Coverage differences:
    - This dataset covers a subset of blockchains but is likely more exhaustive within covered chains.
    - Chainalysis likely has broader blockchain coverage.
    - Chainalysis “indirect” category attempts to trace transfers via self-custodial wallets but likely captures only a fraction of such transactions.

### Aggregate quantity comparison
- Total estimated quantities:
  - This paper’s dataset: $2 trillion
  - Chainalysis: $1.7 trillion
- Agreement between datasets:
  - USDT more prevalent in emerging market (EM) regions; USDC more favored in advanced economy (AE) regions.
  - Asia and Pacific exhibits largest stablecoin flows.
  - Africa and the Middle East, and Latin America and the Caribbean have smaller absolute volumes.
  - Direct (CEX-to-CEX) estimates broadly agree that net stablecoin outflows predominantly originate from North America to other regions.

### Direct vs Indirect category comparison
- Constructed analogous “direct” and “indirect” categories for this dataset:
  - Direct: direct flows between CEXs.
  - Indirect: flows involving self-custodial wallets (CEX↔self-custodial and self-custodial↔self-custodial).
- Findings:
  - Direct category: no significant disagreement between datasets; both show net outflows from North America to other regions.
  - Indirect category: severe disagreement.
    - This paper’s indirect category estimates stablecoins primarily flow from North America to other regions.
    - Chainalysis indirect category suggests stablecoins flow out of all regions into North America.
    - The discrepancy is puzzling since both indirect categories aim to capture similar transfers involving self-custodial wallets and CEXs.
    - The Chainalysis indirect category’s disagreement with other estimates raises questions about its reliability.

### China-specific disagreement
- VPN usage in China likely causes Chainalysis to undercount activity.
- Comparative figures:
  - Gross flows involving China:
    - This paper: $153 billion
    - Chainalysis: $28 billion
    - Ratio: 5.5 times larger in this paper’s estimate.
  - Net flows involving China:
    - This paper: $18 billion
    - Chainalysis: $0.18 billion
    - Ratio: 100 times larger in this paper’s estimate.

*Source: wpiea2025141-source-pdf - 6.3 Dividing Centralized Exchange Flows into Regional Flows*

### 8.2 Correlations

### 8.2 Correlations

### Gross flows correlations between datasets
- Time series constructed: daily stablecoin flows from January 1st 2024 to December 31st 2024, restricted to weekday data only to mitigate weekend cyclicality.
- Very large, positive correlations between outflows of the respective categories in our dataset and Chainalysis, ranging from 0.78 to 0.96 (cf. Table 10).
- Correlations by region (outflows, Our dataset vs Chainalysis):
  - Africa and Middle East: Direct 0.88, Indirect 0.79, Direct + Indirect 0.91
  - Asia and Pacific: Direct 0.89, Indirect 0.85, Direct + Indirect 0.94
  - Europe: Direct 0.90, Indirect 0.89, Direct + Indirect 0.96
  - Latin America and Caribbean: Direct 0.86, Indirect 0.78, Direct + Indirect 0.87
  - North America: Direct 0.91, Indirect 0.88, Direct + Indirect 0.95

### Net flows correlations and divergences
- High, positive correlation for net flows between the direct category in Chainalysis and the direct category in our dataset.
- The indirect category in Chainalysis exhibits correlations with our dataset that range from large negative to negligible, indicating disagreement in direction and dynamics of net flows between datasets (cf. Table 11).
- Correlations by region (net flows, Our dataset vs Chainalysis; format: Direct, Indirect, Direct + Indirect, Chainalysis Direct vs Indirect):
  - Africa and Middle East: 0.27, -0.18, 0.10, -0.42
  - Asia and Pacific: 0.75, 0.01, 0.50, -0.13
  - Europe: 0.22, -0.15, 0.35, -0.27
  - Latin America and Caribbean: 0.60, -0.38, 0.09, -0.47
  - North America: 0.80, -0.22, 0.53, -0.39
- Implication: the Chainalysis indirect category not only disagrees with the direction of net flows in our dataset (and Chainalysis direct), but also with the time-series dynamics.

### Inflow-outflow correlations within datasets
- Prior research noted very high inflow-outflow correlations up to 99% in the Chainalysis dataset.
- Our analysis finds similarly large inflow-outflow correlations to Chainalysis, suggesting this is a salient feature of the underlying data (detailed tables in Appendix: Table16 and Table17).
- In our dataset, computing inflow-outflow correlations purely between self-custodial wallets (not possible in Chainalysis) yields significantly lower correlations than the full-sample inflow-outflow correlations, although they remain high in absolute terms.
  - This suggests that high inflow-outflow correlations are particularly salient for flows involving CEXs, and less so for flows involving self-custodial wallets.
- Inflow correlations reported in the Appendix (Table 16) — Correlation of Inflows between the Respective Categories in Our Dataset and Chainalysis:
  - Africa and Middle East: Direct 0.89, Indirect 0.83, Direct + Indirect 0.93
  - Asia and Pacific: Direct 0.93, Indirect 0.87, Direct + Indirect 0.95
  - Europe: Direct 0.92, Indirect 0.90, Direct + Indirect 0.96
  - Latin America and Caribbean: Direct 0.89, Indirect 0.84, Direct + Indirect 0.91
  - North America: Direct 0.81, Indirect 0.83, Direct + Indirect 0.94
- Inflow-outflow correlations by category for both datasets (Appendix Table 17):
  - Our Dataset vs Chainalysis; entries listed as: Our Dataset Direct, Indirect, Total, Self-custodial; Chainalysis Direct, Indirect, Total
  - Africa and Middle East: 0.99, 0.98, 0.99, 0.89; 0.96, 0.96, 0.98
  - Asia and Pacific: 0.94, 0.98, 0.98, 0.88; 0.97, 0.98, 0.99
  - Europe: 0.99, 0.98, 0.99, 0.84; 0.99, 0.99, 1.00
  - Latin America and Caribbean: 0.99, 0.99, 0.99, 0.91; 0.97, 0.97, 0.98
  - North America: 0.94, 0.93, 0.95, 0.66; 0.94, 0.93, 0.96
- Note: within-region flows were excluded from these calculations to avoid mechanical inflow-outflow correlation.

### Implications for measurement and interpretation
- High gross-flow correlations indicate strong agreement between datasets on magnitude of outflows, but discrepancies in net-flow correlations—especially for the Chainalysis indirect category—highlight important methodological or classification differences that affect inferred directionality and dynamics.
- The marked difference in inflow-outflow correlations between samples involving CEXs and self-custodial wallets implies that exchange-related wallet activity drives much of the high inflow-outflow symmetry observed in aggregate datasets.

*Decrypting Crypto: How to Estimate International Stablecoin Flows — Working Paper No. WP/2025/141*

---


_Source: https://www.imf.org/-/media/files/publications/wp/2025/english/wpiea2025141-source-pdf.pdf_