This slide deck details some of the lessons we learned building price prediction models for cryptocurrencies. The session provides examples and practical tips about the challenges of price predictions in crypto asset markets.
2. Agenda
❖ Some things to know about predictions for crypto
assets
❖ The different approaches
❖ Our journey and lessons learned
❖ Models that work and real time predictions
8. 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
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…
15. 15
Build models that
can forecast short
term price
fluctuations in
crypto-assets
1
Focus on deep
learning methods
2
Start with order
book data source
3
Original Goals
17. Some
Things
We
Learned
17
Crypto orderbook datasets have
many quality issues
Behaviors like wash trading or
spoofing are common
There are many time gaps and
missing data points
Most research papers don’t stand the
test of real market data
Most research methods haven’t been
designed for highly volatile markets
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
23. ● 78 Features
● 52,000 parameters
● 2 LSTM networks trained in the
trade input sequence
● Connects inputs from past and
future bidirectionally
● Ensemble of multiple bi-LSTMs
trained for independent data
sources
23
Bidirectional LSTM
24. Some
Things
We
Learned
24
Crypto orderbook datasets have
many quality issues
Behaviors like wash trading or
spoofing are common
There are many time gaps and
missing data points
Most research papers don’t stand the
test of real market data
Most research methods haven’t been
designed for highly volatile markets
25. Solid
Results
25
Average of 12 predictions
per day
69% accuracy
Retrained periodically
Tested against real time prices
Solid
Results
31. Solid
Results
31
What did
we learn?
Feature engineering matters A LOT!
The more high-quality training data, the better
Periodic retraining is important
No single prediction model beats the market all
the time
Edge cases might require specialized models
Be prepared to fail, A LOT!
32. 32
● ITB will be launching several predictive signals in early Q2. We need your
help to get there!
● Signup for ITB to get an early preview
● Tells us how would you use predictive models (ex: APIs, notifications,
visuals)
● What frequency of predictions matters to you? (ex: hourly, daily?)
● What would you like to see from ITB to TRUST our predictions( ex: real
time accuracy, failure impacts….)
Crypto Market & ITB: We Need Your Help!
33. 33
● Crypto asset predictions are a solvable problem
● No single model can solve the crypto market
● Deep learning models are computationally expensive and hard to
implement but offer an interesting edge over alternatives
● The first version of ITB predictions will be available in early Q2 2020
Summary