REGIME CHANGE AND TREND PREDICTION FOR BITCOIN TIME SERIES DATA

Bitcoin time series dataset recording individual transactions denominated in Euro at the COINBASE market between April 23, 2015 and August 15, 2016 is analyzed. Markov switching model is applied to classify the regions of varying volatility represented by three hidden state regimes using univariate...

Full description

Saved in:
Bibliographic Details
Main Authors: Kodama, Osamu, Pichl, Lukáš, Kaizoji, Taisei
Format: Conference Proceeding
Language:English
Subjects:
Autoregressive models
Autoregressive processes
Computer simulation
Currencies
Digital currencies
Econometrics
Economic models
Markov chains
Neural networks
Recurrent neural networks
Switching theory
Time series
Volatility
Citations: Items that cite this one
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Bitcoin time series dataset recording individual transactions denominated in Euro at the COINBASE market between April 23, 2015 and August 15, 2016 is analyzed. Markov switching model is applied to classify the regions of varying volatility represented by three hidden state regimes using univariate autoregressive model and dependent mixture model. Causality extraction and price prediction of daily BTCEUR exchange rates is performed by means of a recurrent neural network using the standard Elman model. Strong correlations is found between the normalized mean squared error of the Elman network (out-of-sample 5-day-ahead prediction) and the realized volatility (sum of minute returns squared throughout the trading day). The present approach is calibrated using simulated regime change in standard econometric models. Our results clearly demonstrate the applicability of recurrent neural networks to causality extraction even in the case of highly volatile cryptocurrency exchange rate time series data.
ISSN:1805-997X
1805-9961
DOI:10.12955/cbup.v5.954