Deep Learning Algorithms Applied to Blockchain-Based Financial Time Series (FTS).
The main project was approached via multiple sub-tasks or exercises, before building up the final model. All R scripts corresponding to each sub-task can be found in the src directory, with corresponding datasets in the datasets directory. A short PDF document accompanies each exercise to present a summary of results, all of which can be found in the doc directory.
This exercise investigates linear AR Yule-Walker and AR Burg models applied to FTS prediction.
These exercises investigate linear ARMA & ARIMA models applied to FTS prediction and why linear models are not necessarily the most suited for modelling FTS.
This exercise investigates linear ARCH & GARCH models applied to FTS prediction and why linear models are not necessarily the most suited for modelling FTS.
This exercise investigates non-linear Single Layer Perceptron (SLP) & basic Neural Networks (NN) and their performance for modelling FTS.
This exercise investigates Single Hidden Layer Neural Networks (NNetAR algorithm) and their performance for modelling FTS.
This exercise investigates Self Organising Maps (SOM) and their performance in unsupervised learning.
This exercise investigates the differences in applying the AR, ARMA, ARIMA, ARCH, GARCH & SLP models to Forex & BTC FTS (binary only).
This exercise investigates applying the Nonlinear Auto Regressive Moving Average model with eXogenous inputs (NARMAX) to both predicting Forex & BTC FTS (trend & binary).
This exercise investigates Self Organising Maps (SOM) and their performance in unsupervised learning applied to Forex & BTC FTS.
This exercise investigates Echo State Networks (ESN) and their performance in unsupervised learning applied to Forex & BTC FTS.
This exercise investigates Low Frequency Granger Causality in cryptocurrencies.
This exercise investigates High Frequency Granger Causality in cryptocurrencies.
This exercise investigates a preliminary High Frequency NN for forecasting BTC close price with OCLH data.
This exercise investigates a slightly improved High Frequency NN for forecasting DASH close price with delayed time series and exogenous BTC delayed inputs.
This exercise investigates briefly fully Bayesian estimations of stochastic volatility via Markov chain Monte Carlo methods in the BTC-USD pair compared to other typical Forex currencies. For context only.
This exercises looks at the impact of going from HF minute-data to MF-hour data as well as other simple NN architectures.
This exercises looks at Deep NN architectures using Keras, TensorFlow & CUDA.
This exercise investigates Hourly Granger Causality in cryptocurrencies.
This exercises looks at further DNN architectures (up to 20x20 with 2 hidden layers) using Keras, TensorFlow & CUDA.
This proof of concept exercise uses the most performant algorithm developped thus far to trade live on the Poloniex exchange.
These equations cover the bare fundamentals of all the models used. This can also be found in the doc directory.
(Complete, graded 80.50%)
(Complete, graded 92.00% overall)