This document discusses predictions in crypto assets using deep learning. It outlines different data sources that can be used for predictions, including order book data, blockchain data, and alternative datasets. Two main approaches to predictions are described: time series models that use linear correlations, and machine learning models using neural networks. Examples are given of convolutional and LSTM neural networks that could be used to build prediction models based on order book data. The document also discusses some myths around predictions, such as the idea that a single attribute or model can predict prices across all conditions. It describes the work of IntoTheBlock to build various deep learning models for crypto predictions using different datasets.
2. Agenda
❖ A gentle introduction to predictions in capital markets
❖ Prediction In crypto assets
❖ Myths and realities of crypto asset predictions
❖ Time series vs. deep learning architectures
❖ Some ideas to explore
7. Asset-Based
Predictions
Ex: Predict the price of Bitcoin
in the next 12 hours
Focus on predicting the
performance of a single asset
based on a specific criteria
Typically factors in specific
characteristics of the target
asset
8. 8
Asset-Based Predictions
Pros
• Rich models based on the specific characteristics of an asset
• Possible to achieve high levels of accuracy during short periods of
time
• Simpler to retrain and validate
Cons
• Exposure to a single asset within a market
• Hard to assemble in cohesive portfolios
• Becomes challenging to combine different predictions as part of
portfolios
9. Factor-Based
Predictions
Ex: Predict is cryptos with strong
momentum will outperform in
the next 24 hours
Focus on predicting the
performance of a group of assets
based on a set of factors
Typical factors include aspects
such as value, momentum, carry,
volatility, quality, etc…
10. 10
Factor-Based Predictions
Pros
• Minimum exposure to a single asset within a market
• Ability to formulate sophisticated thesis across an entire portfolio
• Exposure to non-correlated factors and hedging models
Cons
• Prediction models need to ignore unique characteristics of a
given asset
• Prediction models might miss unique investment opportunities in
a specific asset within the portfolio
14. Predictions Based on Order Book Datasets
14
Focus on predicting short term performance
fluctuations for an asset or a factor
Leverage attributes such as price, volume as well as
order book bids, asks and trades
Models are vulnerable to the poor quality of crypto
exchange order book datasets
15. Predictions Based on Blockchain Datasets
15
Focus on predicting medium to long term performance
trends on a crypto-asset or a factor
Leverage attributes such as hash rates, exchange flows,
wallet balances, etc.
Models are very unique to crypto assets and can’t take
advantage of existing quant research
16. Predictions Based on Derivative Contracts
16
Focus on predicting the short or medium-term
impact of derivative contracts in crypto markets
Current prediction models would necessarily be
asset based
Unique insights when combined with blockchain
datasets
17. Predictions Based on Protocol Contracts
17
Focus on medium to long term predictions of the
performance of specific protocol
By definition, prediction models will be asset based
Vulnerable to unexpected changes and versions of
a given protocol
18. Predictions Based on Alternative Asset Datasets
18
Focus on short term predictions based on sources like
Twitter, Telegram, News, etc.
Leverage natural language understanding techniques to
extract information from text data sources.
Largely used as a complement to more sophisticated
prediction models
19.
20. 20
The Super Predictor Myth
Myth
• A single attribute can predict the price of a crypto
asset
• Ex: NVT score can predict the price of Bitcoin….
Reality
• Most prediction models require a combination of
many factors
21. 21
The Super Model Myth
Myth
• A single model can predict the price of a crypto asset
across different market conditions
• Ex: A model based on technical that can regularly predict
the price of Bitcoin
Reality
• Most models only work under specific market conditions
and they tend to overfit for those
22. 22
The Price-Volume Myth
Myth
• All characteristics of a crypto asset are captured in
price and volume sentiment
Reality
• Blockchchain and derivatives datasets offer unique
insights about crypto assets that are impossible to
capture in price and volume models
24. Time Series
Prediction
Models
24
Statistical models that focus on predicting a variable
based on linear correlations with other attributes of
the dataset
The value of the predictor attributes in the future
should be known in advance
Most algorithms are based on variation of linear
regressions
Ex: predict the price of Bitcoin based on the expected
settlements in Bitcoin futures
Algorithms: ARIMA, ARIMAX, VARMAX
25. Machine
Learning
Prediction
Models
25
Neural network architectures that forecast a variable
based on hierarchical, non-linear relationships
between other predictors
Analyze complex relationships in sophisticated
datasets
Different schools of thought: supervised,
unsupervised, semi-supervised, reinforcement, etc
More prompt to discover alpha in efficient markets
Proven algorithms for financial markets: long-short-
term memory networks, convolutional neural
networks, multi-layer perceptrons
29. Example: CNN+LSTM
for Order Book Data
• The input is an image of the order
book from multiple exchanges
• A CNN layers detecting features for
the order book data
• A time distributed layer identifies
patterns for a specific timeframe
• A LSTM layer that aggregates all the
individual patterns into a target
prediction
29
30. Example: Multi-Layer
Perceptron for Order
Book Data
• The input is based on trades in the order
book of different exchanges
• A layer that obtain connections between
different features
• A series of layers that detect more
sophisticated patters
• Another layer that limits potential
overfitting error
• A prediction layer that aggregates the
patterns into potential predicitons
31. 31
Summary
● Crypto assets offer a blank canvas for prediction models
● There are different data sources of predictions for crypto assets: order
book, futures, blockchain, alternative …
● The two main schools of thought for prediction are : time series and
machine/deep learning
● ITB is working on prediction models for crypto assets based on deep
learning