ABSTRACT
Despite the growing literature on Bitcoin and other cryptocurrencies, we know relatively little about who are involved in trading, transacting and using these assets and how they behave. Examining millions of Bitcoin transaction records, we show that less than 1% of Bitcoin users contribute to more than 95% of the market volumes. These ‘whales’ are often associated with strategic trading/transaction volumes, market reactions and timing patterns. Using K-means clustering on a comprehensive transaction dataset, we establish a typology of traders by learning their trading exchange patterns, strategies and impact risk and market microstructure. Our approach ‘learns’ and identifies five distinct groups or types of Bitcoin users, which are somewhat, though not entirely, comparable to popular categorisations used in conventional market such as fundamental, technical, retail and institutional traders as well as market makers. Four of these groups present distinguishable trading patterns with a strong impact on liquidity provision and trading signals.
RngIntvi (cid:8) MeanIntvi)2 N(i)1 MeanIntvi (i) N(i) D D H(i) k 1>0((cid:3)(i) H(i) (cid:8) k 1>0((cid:3)(i) H(i)/7 k 1>0((cid:3)(i) H(i)/30 j StdIntvi = (i) 1 ActTime2Li = 100 (cid:8) ) TransD2Li = 100 D,k ) W,k TransW2Li = 100 (cid:8) ) M,k TransM2Li = 100 Table 2. Summary of Bitcoin trader types. Class Num. of addr. Trader category Matches to stereotypes 1 2 3 4 5 2, 089, 694 2, 101, 244 1, 023, 726 566 269 Bitcoin tasters Liquidity takers Liquidity providers NA Fundamental traders Technical traders Market makers High-frequency traders
id: f165598bd3bb0d28dbad9da14b0884e1 - page: 9
4.2. Trader classes and their conventional parallels As the K-means algorithm starts with a randomized classification, the results may vary when using different random seeds. To observe robust components of each trader class, we run the algorithm 10 times and collect consistent classifications. Finally, we find that 5,215,499 (out of 6,108,128) addresses have the same classification throughout all runs. According to the centroid vectors of K-means clustering (see Table 3), we find some clear characteristics of traders in each class and are able to associate some of them with typical types of traders in the traditional financial market (summarized in Table 2). Note that we name the trader classes based on the centroid.
id: d0b57e63cf989c9925112739f31a50c9 - page: 9
The Class #1 centroid indicates that this class is, in general, not active. Traders in this class only contribute a few transactions 1 transaction per block and 2 transactions per active day (see Table 3(a)); only having exchanges/transactions in the first 20.07% of time after joining in the market (see Table 3(c)). The deviation of THE EUROPEAN JOURNAL OF FINANCE Figure 3. WSSE for K selection. Table 3. K-means clustering results. (a) Trading frequency features Class CntTransB CntTransD CntTransW StdCntTransW 1 2 3 4 5 1.07 1.04 1.05 2.08 13.95 2.00 1.71 1.85 70.38 560.35 4.28 3.01 4.66 389.43 2569.68 0.53 0.49 0.57 0.82 0.85 (b) Waiting time features Class MedIntv (Days) MeanIntv (Days) RngIntv StdIntv 1 2 3 4 5 1.57 3.21 2.24 0.02 0.01 7.27 22.36 4.86 0.14 0.11 8.92 13.72 13.04 1893.16 366.16 1.87 2.33 1.74 27.01 8.89 (c) Active trading features Class ActTime2L (%) TransD2L (%) TransW2L (%)
id: 266ba0b1e0139b29e9c45164ffe09755 - page: 9
TransM2L (%) 1 2 3 4 5 20.07 72.65 88.37 72.35 31.24 3.07 4.63 21.24 34.13 24.61 10.24 17.84 61.90 40.68 28.04 19.38 39.56 88.53 48.11 33.05 waiting time is small, e.g. RngIntv is 8.92, which also tells us that transactions occur with a consistent rhythm, e.g. one trade per day, without an indication of information-based trade decisions. Traders in this class are not really engaging in Bitcoin trading and transactions and behave like tasters. We call this a casual trader class. 9 10 A. LIU ET AL. Figure 4. Class #2 behavioural specics. (a) Active blocks vs. Active days and (b) Number of transactions vs. Active blocks.
id: b4a6842a0f58fe10daeb5d44c8579c87 - page: 10