Show me, don’t tell me should be the mantra when it comes to cryptocurrency predictions! A lot of written research but no real world implementations. At IntoTheBlock, we’ve spent months working on predictive models for crypto-assets. We have failed a lot and we have learned a lot. So it is time to show some results!
In this session, we will show you deep neural network architectures that are predicting the price of crypto-assets in real-time based on different goals. We will share some of the lessons we learned trying to predict recent market events and some of the new ideas we are working on. We will showcase different predictive models and reveal some of their most recent predictions.
Find out more at https://app.intotheblock.com/
2. Agenda
● IntoTheBlock’s journey building
predictive models for crypto assets
● Practical lessons:
○ What to predict?
○ How to predict it?
○ How to operationalize it?
● What we learned
● Some results
5. In-Depth Analysis
Deep neural networks can
uncover new opportunities in
alpha generating strategies
Resiliency
Emerging deep learning models
have proven to be more resilient
to market changes than time
series or basic machine learning
models
Defensibility
Deep neural networks are harder
to trick and recreate than
alternatives
A Deep Learning Approach
9. Partnering with several quant
funds on the implementation of
trading strategies based on the
output predictive models
Collaborating with research
groups in new models for crypto
asset price predictions
Launching a version of prediction
services for retail investors in the
IntoTheBlock platform
Getting Real
19. ● ARIMA, DeepAR+, Prophet
● Easy to implement and fast to execute
● Poor resiliency to market fluctuations
● Limited number of potential predictors
● Hard to estimate predictors ahead of time
19
Time Series Forecasting Models
20. ● Linear regression, decision trees
● There is a lot of research available in
this area
● Poor resiliency to market
fluctuations
● Hard to achieve knowledge
generalization(underfitting)
● Prompt to overfit
20
Traditional Machine Learning Models
21. ● Computationally expensive to
execute at scale
● Difficult to interpret
● Many of the benefits such as
automated feature extraction are
hard to materialize
● Great to tackle sophisticated theses
21
Deep Learning Models
22. The case for deep learning methods in
predictive strategies…
23. They can uncover
complex non-linear
relationships between
variables
Large body of Research Resiliency
Three Things to Consider About Deep Learning in
Predictive Models
1 2 3
24. The near-term future of predictive models
will be driven by deep learning methods…
25. Let’s look at a few examples of deep
learning innovations for crypto assets…
26. ● Deciding which attributes are predictors of price
● Ex: Are order books’ bid-ask predictors of price?
26
Challenge: Feature Extraction
27. Convolutional
Neural Networks
• Extract relevant features
from an input image
• Its relevant to crypto to
generate automated
features from a dataset
such an orderbook
28. ● Deciding which architecture adapts better to a given predictive problem
● Ex: Which deep learning method to for predicting price based on Bitcoin
trades?
28
Challenge: Model Selection
29. Neural Architecture
Search
• Uses machine learning to
test different machine
learning models for a
given problem.
• Its relevant to crypto to
discover sophisticated
models for specific tasks.
30. ● Building resiliency to missing or inaccurate data about crypto assets
● Ex: Bitcoin order books in centralized exchanges often include data gaps
or fake trades
30
Challenge: Data Accuracy / Gaps
31. Semi-Supervised
Learning
• Enables the training of
models with minimum
datasets
• Its relevant to crypto
because the lack of
labeled datasets for
crypto-assets.
32. ● Reusing knowledge acquired in a crypto-asset prediction task in other
tasks
● Ex: After building a predictive model for Bitcoin prices using order books,
can we build a similar model for Ethereum?
32
Challenge: Knowledge Reusability
33. Transfer
Learning
• Allows to reuse models
across different tasks
• Its relevant to crypto in
order to build models that
reuse other trained
models.
34. ● Train a model for unforeseen market conditions in the absence of data
● Ex: Can I train a model to be resilient to abrupt Bitcoin price crashes
before they happen?
34
Challenge: Training Coverage and Acceleration
35. Generative
Models
• Create new data
instances based on
previous training models
• Its relevant to crypto to
generate new datasets of
simulated market
conditions.
58. New Predictive Models
Order Books
5-10 predictions based on
trade information
1-2 hour predictions based
on orders
Derivatives
Predict short term 1-2 hours
price movements based on
futures activity
Blockchain
Predict 5-10 hours predictions
based on exchange flows
1 2 3
59. 59
● Crypto assets show predictable capabilities during different market times
● Building a predictive models requires formulating a thesis, selecting a
methodology and building the required infrastructures
● Deep learning offers interesting choices for crypto asset price predictions
● Signup at http://app.intotheblock.com prior to our predictions release
Summary