货币主权面临稳定币带来的新挑战

Decrypting Crypto: Howto Estimate International Marco Reuter WP/25/141 IMF Working Papersdescribe research inprogress by the author(s) and are published toelicit comments and to encourage debate.The views expressed in IMF Working Papers are 2025JUL IMF Working PaperResearch Department Decrypting Crypto: How to Estimate International Stablecoin Flows Authorized for distribution by Maria Soledad Martinez Peria IMF Working Papersdescribe research in progress by the author(s) and are published to elicitcomments and to encourage debate.The views expressed in IMF Working Papers are those of the ABSTRACT:This paper presents a novel methodology—leveraging a combination of AI and machine learningto estimate the geographic distribution of international stablecoin flows, overcoming the “anonymity” of cryptoassets. Analyzing 2024 stablecoin transactions totaling $2 trillion, our findings show: (i) stablecoin flows arehighest in North America ($633bn) and in Asia and Pacific ($519bn). (ii) Relative to GDP, they are most significantin Latin America and the Caribbean (7.7%), and in Africa and the Middle East (6.7%). (iii) North America exhibits WORKING PAPERS Decrypting Crypto: How toEstimate International Stablecoin Prepared by Marco Reuter 1Introduction Policymakers are increasingly wary of the popularity of crypto assets and have called forbetter monitoring of crypto transactions and international crypto asset flows (BIS (2023), EU (2023), G7 (2023) , FATF (2023), FSB (2023), IMF (2023), US Treasury (2023)). At thesame time, recent research shows that crypto assets are increasingly used for internationaltransactions, particularly when capital flow measures make it difficult to use traditional The main contribution of this paper is the development of a novel method that enablesthe identification of the geographic regional origin of crypto wallets, facilitating the measure- ment of international stablecoin flows.1Before detailing our method, we address a commonmisconception—contrary to popular belief, the vast majority of crypto assets donotprovideanonymity. Every transaction is publicly recorded on a freely accessible ledger known as a '0xdFDEe1155E1dd7c01774560C6E98C41B7da945dB', does not directly reveal personal in-formation about the user. The key challenge in mapping the geography of crypto asset flowsis supplementing blockchain data with useful information about senders and receivers. Ourmethodology addresses this challenge by enabling the estimation of the geographic region ofany arbitraryself-custodial wallet2in the Ethereum ecosystem. To estimate the geographic region of self-custodial wallets (we assign wallets to one ofthe following five regions: Africa and the Middle East, Asia and the Pacific, Europe, NorthAmerica, and Latin America and the Caribbean), our methodology involves obtaining ge- ographic information for a subset of wallets through two distinct approaches.First, we markers—such as language, script, or regional references—that suggest a wallet’s likely re-gion. Second, we identify wallets that frequently transact with centralized exchanges (CEXs)targeting specific regional markets, assuming that a wallet predominantly interacting with, The core of our approach lies in leveraging this training data to train a machine learningmodel to recognize patterns in on-chain activity that are indicative of a wallet’s geographicorigin. We construct features capturing wallets’ behavioral and transactional characteristics,including time-of-day activity patterns, adherence to daylight savings time, interactions withcertain centralized exchanges, and engagement with popular ERC-20 tokens and smart con- Using this approach, we analyze almost 6 million domain names and billions of on-chain transactions to construct the training dataset. Some examples of regions assigned byanalyzing domains names with the use of a LLM are, “pijiu” (Chinese for “beer”) which isassigned to Asia and the Pacific, andﺍﻝﻙﻱﻡﺍﻭﻱﺍﺕ(Arabic for “chemicals”) which is assignedto Africa and Middle East. We validate these ad hoc classifications through time-of-day basedactivity profile analysis. By adjusting wallet transaction timestamps to regional time zones We then train a Gradient Boosted Decision Tree model on this dataset, which achievesan overall accuracy of 65% in predicting geographic regions. For context, random guessingwith five regions yields 20% accuracy.We then apply the model to predict the region of We document significant regional variation in stablecoin usage.In absolute terms, weestimate that Asia and the Pacific lead with the highest stablecoin activity (inflows: $407bn,outflows: $395bn, intraregional flows: $209bn), followed by North America (inflows: $363bn,outflows: $417bn, intraregional flows: $216bn). However, relative to GDP, Africa and theMiddle East, and Latin America and the Caribbean stand out, with stablecoin usage reach-ing 6.7% and 7.7% of GDP, respectively.Additionally, intraregional flows i