You are on page 1of 46

Get your head in the game

CRYPTO
DECRYPTO!
Unlocking the Secrets of Bitcoin with Neural Networks

Presented by LT5: Aguila | Amorado | Ang | Licong


Bitcoin

01 Background
Cryptocurrency invented in 2008

02 Bitcoin Experience
Mining, Wallets, Investments
Winner of the hearts of many
8000x increase from 2012

Bitcoin Price
Bitcoin Price chart
Bitcoin Price chart
Bitcoin Price chart
Bitcoin Price chart
Bitcoin Price chart
Problem Statement

The million dollar question

Can we consistently generate profits from


trading the daily swings of Bitcoin?
The Dataset

Source: Kaggle

Bitcoin 1-minute data from 2012 to 2020


~4.7 million rows
Methodology
How we did it
Methodology
How we did it

Preprocessing
Aggregation from 1-minute to hourly
Feature Engineering (e.g. moving averages)
Methodology
How we did it

Preprocessing Splitting the Data


Aggregation from 1-minute to hourly Train on 2013-2017 data
Feature Engineering (e.g. moving averages) Validate using 2018-2019, Test on 2020
Methodology
How we did it

Preprocessing Splitting the Data Neural Network


Aggregation from 1-minute to hourly Train on 2013-2017 data Model tuning and evaluation
Feature Engineering (e.g. moving averages) Validate using 2018-2019, Test on 2020
Model Generation
What makes the Bitcoin prediction go brrrrr

Feedforward ~900 parameters


Neural Network
15 minutes training
2 hidden layers time on GPU

~70 unique models ran


Model Performance
Goal: Go long, short, or do nothing in a 24-hour forward window
Model Performance
Goal: Go long, short, or do nothing in a 24-hour forward window

48%
Accuracy
Model Performance
Goal: Go long, short, or do nothing in a 24-hour forward window

48% 44%
Accuracy Precision
Model Performance
Goal: Go long, short, or do nothing in a 24-hour forward window

48% 44% 5%
Accuracy Precision Recall
Backtesting
Backtesting
Model

Buy
and hold
Backtesting
Model

Buy
and hold
10,000 USD
Backtesting
41,760 USD Model

Buy
and hold
10,000 USD
Backtesting
Model: 318% in 3Y = 61% CAGR 41,760 USD Model
Buy-and-hold: 112% in 3Y = 28% CAGR
Buy
and hold
10,000 USD
Backtesting
Model: 318% in 3Y = 61% CAGR 41,760 USD Model
Buy-and-hold: 112% in 3Y = 28% CAGR
Buy
21,164 USD
and hold
10,000 USD
Backtesting
Model: 318% in 3Y = 61% CAGR 41,760 USD Model
Buy-and-hold: 112% in 3Y = 28% CAGR
Buy
21,164 USD
and hold
10,000 USD
Backtesting
Model: 318% in 3Y = 61% CAGR 41,760 USD Model
Buy-and-hold: 112% in 3Y = 28% CAGR
Buy
21,164 USD
and hold
10,000 USD
Backtesting
Model: 318% in 3Y = 61% CAGR 41,760 USD Model
Buy-and-hold: 112% in 3Y = 28% CAGR
Buy
21,164 USD
and hold
10,000 USD
11/17

Key Trade Statistics


Number of Trades Number of Longs Number of Shorts

152 106 46

Hit Ratio Ave Gain/Trade (net)

54.6% 0.94%
9/17

Conclusion

Secrets
Unlocked!
THANK YOU!
We hope you learned something new!

Visit us later @ tinyurl.com/crypto-decrypto


Q&A
Features Used (30 total)
Returns. 1-hr, 2-hr, 4-hr, 12-hr, 24-hr, 48-hr, 96-hr, 192-hr change

Distance to moving averages. 10-MA, 20-MA, 50-MA, 100-MA, 200-MA

Slope of moving averages. 10-MA, 20-MA, 50-MA, 100-MA, 200-MA

Momentum indicators. RSI 14, RSI 30

Volume indicators. 24-hr MA, 96-hr MA, ratio of 24/96 MA

Volatility indicators. 24-hr stdev, 96-hr stdev, ratio of 24/96 stdev, 12-hr
drawdown, 24-hr drawdown, 48-hr drawdown, 96-hr drawdown
Sample Neural Network
Architecture

30 features 20 nodes 12 nodes 3 outputs


Number of Trades Made
Maximum Drawdown
Trade Logs (2018)
Trade Logs (2019)
Trade Logs (2020)
Binance Fee Schedule
Source: https://www.binance.com/en/fee/schedule
Cryptocurrency
Cryptocurrency
Cryptocurrency
Bitcoin Mining
Source: https://medium.com/@PhillipNunn/the-bitcoin-supply-curve-513554e6588d
Technical Analysis Principles
Contact us!

Nico Aguila AlphonsoAguila.MSDS2021@aim.edu

Adriane Amorado AdrianeAmorado.MSDS2021@aim.edu

Perry Ang PerryAng.MSDS2021@aim.edu

Claudine Licong ClaudineLicong.MSDS2021@aim.edu

You might also like