Enhancing Bitcoin Price Volatility Estimator Predictions: A Four-Step Methodological Approach Utilizing Elastic Net Regression

Georgia Zournatzidou, IOANNIS MALLIDIS, Dimitrios Farazakis, Christos Floros

Keywords: elastic net regression, volatility estimators, time-series analysis

TL;DR: A four-step Machine Learning approach for predicting Bitcoin price volatility estimators using Elastic Net Regression and Random Forest Regression

Abstract:

This paper provides a computationally efficient and novel four-step methodological approach
for predicting volatility estimators derived from bitcoin prices. In the first step, open, high,
low, and close bitcoin prices are transformed into volatility estimators using Brownian motion assumptions
and logarithmic transformations. The second step determines the optimal number of
time-series lags required for converting the series into an autoregressive model. This selection process
utilizes random forest regression, evaluating the importance of each lag using the Mean Decrease
in Impurity (MDI) criterion and optimizing the number of lags considering an 85% cumulative
importance threshold. The third step of the developed methodological approach fits the Elastic
Net Regression (ENR) to the volatility estimator’s dataset, while the final fourth step assesses the
predictive accuracy of ENR, compared to decision tree (DTR), random forest (RFR), and support
vector regression (SVR). The results reveal that the ENR prevails in its predictive accuracy for open
and close prices, as these prices may be linear and less susceptible to sudden, non-linear shifts
typically seen during trading hours. On the other hand, SVR prevails for high and low prices as these
prices often experience spikes and drops driven by transient news and intra-day market sentiments,
forming complex patterns that do not align well with linear modelling.

Serve As Reviewer: ~Dimitrios_Farazakis1

Submission Number: 19