"Lessons Learned from running a quant crypto fund" presented by Michael Feng, CEO and Co-founder of hummingbot
1. Crypto enables new quant strategies
2. Build a chain of production
3. Preventing overfitting is job #1
4. Establish a disaster response plan
5. Every model has an expiration date
Learn more about algo crypto trading: https://www.hummingbot.io
4. CoinAlpha
Mission: To give individuals financial superpowers
using blockchain technology
Deep domain expertise in trading, corporate finance,
software engineering, and data science
Based in Mountain View, CA
5.
6. Our hedge fund
CoinAlpha Falcon, LP
Oct 2017 to Sept 2018
BTC/USD, ETH/USD
Directional (momentum, mean reversion)
Mid-frequency (1-3 days)
22 ($600k)
1% base / 20% performance
None (Ethereum smart contract)
Name
Track record
Portfolio
Strategy
Frequency
Investors
Fees
Administrator
7. Fund performance
Falcon BTC ETH
Return 21.8% 13.3% -28.9%
Sharpe ratio 0.5 0.2 -0.3
Sortino ratio 0.74 0.27 -0.50
Max drawdown -40.5% -68.5% -83.2%
8. Our new product
High frequency market making bot
Open source
Professional-grade
Cross-exchange market making
DEX-compatible
hummingbot.io
11. Inefficient market
Hundreds of exchanges globally
New instruments appear daily
No standard FIX protocol
Lack of large institutional players
Visible trends in historical data
15. 1. Crypto enables new quant strategies
Directional
momentum
mean reversion
correlation
sentiment
event-driven
Overfitting
Arbitrage
jurisdiction
spot vs futures
CEX vs DEX
borrow/lend
Saturation
Market making
liquid, high volume assets
Illiquid, low volume assets
cross-exchange
Inventory
Type
Variations
Primary risk
16. 2. Build a chain of production
Data
curation
Feature
analysis
Strategy
definition
Backtesting Deployment Monitoring
Collect, clean,
index, store data
Discover features
by labeling and
weighting data
Develop a
general model
that explains the
features
Assess the
profitability of a
model using
historical data
Roll out the
strategy into live
trading
Evaluate the
performance of all
live strategies and
allocate capital
between them
17. 3. Preventing overfitting is job #1
Why it happens
● Humans naturally find patterns
● Time-series data is not independent
● When in doubt, add more features!
How to prevent it
● Separate tasks
● Split data into training/validation/test
● Paper trade before doing live trading
● Be skeptical!
18. 4. Establish a disaster response plan
Falcon ETH trading Dec 15-22, 2017
1. Bot buys
2. Manual intervention
3. Bottom
4. Bot sells
19. 5. Every model has an expiration date
Market regime pre May 2018
Trends last for days/weeks
Higher volatility
Long term up and down cycles
Market regime post May 2018
Trends last for hours/days
Lower volatility
Long-term downtrend
21. Books
Advances in Financial Machine Learning by Martin Lopez de Prado
Flash Boys by Michael Lewis
The Quants by Scott Patterson
When Genius Failed by Roger Lowenstein
22. Courses
Machine Learning (Coursera, free)
Intro to Machine Learning for Coders (fast.ai, free)
Artificial Intelligence for Trading (Udacity, paid)
Cryptocurrency Trading (Blockchain at Berkeley, free)
23. Open source projects
hummingbot: high frequency market making bot
ccxt: trading API for multiple crypto exchanges
gekko: TA-based trading bot and backtesting tool
backtrader: Python backtesting library
catalyst: Crypto fork of Quantopian Zipline