Centralized exchanges are one of the most obscure and difficult to understand elements in the crypto landscape. From fake volumes to transaction transformations, centralized exchanges introduce a level of obfuscation that challenges even the most sophisticated analytic techniques. How can we learn to identify and understand the behavior of centralized crypto exchanges?
This session showcases a series of machine learning and data visualization techniques that help us better understand some of the patterns of crypto exchanges. Using gorgeous data visualizations, we will walk you through a journey that clearly illustrates how exchanges process transactions and distribute crypto-assets across their different addresses. Finally, we will illustrate how certain behaviors of crypto exchanges become relevant to specific patterns in the crypto market.
2. Agenda
2
âť– The challenges of centralized exchanges for crypto
analytics
âť– Why understanding centralized exchanges is a machine
learning problem?
âť– What machine learning can tell us about centralized
exchanges ?
4. Centralized Exchanges
4
âť– Are one of the most important components of the crypto
ecosystem
âť– Represent a significant portion of the daily trading volume
âť– Exchange inflows and outflows represent an important
metric of the behavior of the crypto ecosystem
6. Blockchain Analytics Nightmare
6
âť– Exchange trades are not always recorded in public
blockchains
âť– Volume manipulations
âť– Wallet anonymity
âť– Transaction composition patterns
8. Used to
transfer
funds out
of the
exchange
Used to
transfer
funds into
an
exchange
Used to secure
the storage of
crypto-assets.
The main
interaction point
between
external parties
and an
exchange.
Hot
Wallets
Cold
Wallets
Withdrawal
Addresses
Deposit
Addresses
16. Ensemble Learning
• A machine learning technique in
which multiple learners try to
master the same task and are
combined in a single master
model
• Ensemble models in machine
learning combine the decisions
from multiple models to
improve the overall
performance
• Work efficiently on diverse
datasets
18. Inside
IntoTheBlock’s
Exchange Classifier
ITB Exchange
Intelligence
Obtain
features from
the main
actors in
Ethereum
blockchain Create
ensemble
of models
Train on the
Ethereum
blockchain
Predict and
evaluate the
ensemble in the
Ethereum
blockchainGeneralize
the ensemble
to other
blockchains
Test and predict
on non-Ethereum
blockchains
Add blockchain
specific
heuristics
Optimize
model
19. What we Learned by Analyzing Crypto
Exchanges Using Machine Learning and
Fascinating Data Visualizations
39. 3940
Some Challenges Worth Mentioning
âť– Continuous Training: There is not enough labeled datasets about
centralized exchanges for most cryptocurrencies.
âť– Architecture Changes: Exchanges regularly change their transaction
processing patterns.
âť– Privacy/Encryption: Blockchains like Monero or zCash obfuscate information
related to exchanges.
40. 4041
Summary
âť– ML First: Analyzing crypto exchanges is a machine learning problem
âť– More is Better: Combining many machine learning models into ensembles is
an efficient technique to tackle the classification of centralized exchanges
âť– The Core Architecture: The are four key components of centralized
exchanges: hot wallets, cold wallets, deposit addresses and withdrawal
addresses.
âť– Lots of Insights: The analysis of centralized exchanges reveals fascinating
insights that can provide an edge to researchers and traders in crypto
markets