Quant strategies for crypto seem like a perfect match but does it really work? The reality is that most quant strategies for crypto asset fail. This is particularly true for strategies based in machine learning(ML) methods. But why? From data quality to market inefficiencies, the causes are many.
This session explores the fundamental causes of failure of ML-based quant strategies for crypto assets. We will illustrate each cause in the context of real world examples of crypto trades. Additionally, we will explore new ML techniques that can help to improve ML-based quant strategies in the crypto space.
5. 5
● Crypto vs. Capital Markets Insights
--BTC vs. Gold, Commodities, Market Indices, FAANG, TSLA,
Renewable Energy stocks, etc.
● ITB Resources Portal
● New features to be launched in the next few weeks:
--Social sharing
--Data downloads
--New DeFi protocol analytics: Balancer, Aave
--New UI improvements
Some Other Updates
8. 8
Important Clarifications
We are not saying that….
• All crypto quant ML strategies
fail
• Crypto quant models are
doomed
• ML models don’t work in crypto
• ….
17. 17
Crypto is Different
●Datasets are really small compared to traditional
capital markets
Most ML quant models struggle learning with small
datasets
The availability of existing high quality datasets is
very unstable
18. 18
Some Ideas to Consider
Consider ML techniques that work with small datasets: semi-
supervised learning, self-supervised learning
Synthetic training dataset creation using generative models
Build crypto-specific quant models
20. 20
Crypto is Different
●Outlier events happen regularly(ex: Bitcoin recent
rally, March crash, the raise of DeFi…)
ML quant models struggle to learn patterns they
haven’t seen in training datasets
Some outlier events introduce noise in training
datasets
21. 21
Some Ideas to Consider
Rely on ML techniques that work-well with uncertainty. Ex:
reinforcement learning
Build “uncertainty-awareness” in ML crypto quant strategies as a
risk mitigation approach
Retrain and version models regularly
23. 23
Crypto is Different
●Small datasets force ML crypto quant models to optimize
for the training data
Constantly changing market conditions force ML crypto
quant models to regularly incorporate new training data
Overfitted ML crypto quant models fail under real market
conditions
24. 24
Some Ideas to Consider
Develop robust back and live testing mechanisms to detect
overfitting behavior (ex: early stopping, regularization…)
Be strict about detecting bias and variance in training datasets
Develop versioning and rollback mechanisms for crypto quant
strategies
26. 26
Crypto is Different
●Outlier events cause ML crypto quant strategies to fail and
force them to incorporate new records in the training dataset
Training ML crypto quant models with the latest data is a
recipe for overfitting
Outlier training records could distort the behavior of crypto
quant strategies
27. 27
Some Ideas to Consider
Training shuffling methods can help to mitigate the impact of
outlier events
Always maintain multiple versions of crypto quant ML strategies
after retraining
Consider the following three factors for retraining: performance
decay, outlier events and time
29. 29
Crypto is Different
●Crypto datasets are plagued with “bad data”(ex:
wash trades, fake volume…)
Crypto data providers are unstable and fail all the
time
ML crypto quant models fail with bad or missing data
30. 30
Some Ideas to Consider
Rely solely on datasets from trusted exchanges
Design strategies to stop ML crypto quant models in the absence
of real time data
Design backend processes to curate and detect anomalies in
training datasets
32. 32
Crypto is Different
●Anonymous blockchain records introduce a misleading level
of uniformity in training datasets in ML crypto quant models
There is no consistency in the way crypto exchanges
interact with the underlying blockchains
There is no linear correlation between blockchain activity and
trading activity
33. 33
Some Ideas to Consider
Invest in ML classifiers that deanonymize blockchain records
Avoid building ML crypto quant models that rely solely in
anonymous blockchain addresses and transactions
Rely on crypto ML models that incorporate domain knowledge
and semantics based on deanonymized blockchain records
35. 35
Crypto is Different
●Factor investing has been one of the cornerstones
of quant finance
Most traditional factors have no equivalence in
crypto assets
Quant factor strategies can’t be easily adapted to the
crypto space
36. 36
Some Ideas to Consider
Rely on traditional factors more as a risk-hedging mechanisms
Create crypto-specific factors (ex: DeFi momentum, carry etc.)
Avoid trying to adapt traditional factor strategies to the crypto
space
38. 38
Crypto is Different
●Crypto is too inefficient for simple ML quant
strategies
Its hard to model the behavior of crypto assets with
a handful of parameters
Quant research based on simple ML models results
impractical in the crypto space
39. 39
Some Ideas to Consider
Consider deep learning models as the cornerstone of crypto
quant strategies
Invest in interpretability and explainability tools
Keep simple ML quant models to be constrained scenarios
41. 41
Crypto is Different
●Quant infrastructures in crypto are relatively
nascent
Areas such as portfolio management, risk
mitigations need to be adapted to the crypto space
Data and ML infrastructure in crypto quant funds
remains highly experimental
42. 42
Some Ideas to Consider
Invest in data and ML infrastructure for crypto quant models
Leverage open source quant frameworks for back testing,
simulations etc.
Better infrastructures beats better quant strategies over time
44. 44
Crypto is Different
●The crypto space still hasn’t attracted large quantities of
quant talent
Most quant experts still don’t understand crypto and most
crypto traders are not very good at quant strategies
Quant hedge funds in crypto remain relatively small
45. 45
Some Ideas to Consider
ML experts and mathematicians make better crypto quants than
finance people
46. 46
● Crypto was born in the golden era of ML and quant finance
● There are many ML open source frameworks and platforms that can be
adapted to the quant crypto space
● Crypto is the richest asset class in history and that opens the opportunity
for a new generation of quant ML strategies
Despite the Challenges….
47. 47
● Quant ML strategies are definitely the future of crypto trading
● Most traditional ML quant techniques don’t apply in the crypto space
● There are very tangible challenges to build effective crypto quant ML
strategies
● Most crypto quant ML strategies fail but the challenge is worth pursuing
Summary