Gadirli, Farid
(2025)
Cryptocurrency returns, volatility and investor attention: an empirical analysis using econometric and machine learning techniques.
PhD thesis, University of Nottingham.
Abstract
This thesis presents a comprehensive empirical investigation into the rapidly evolving cryptocurrency market, employing a range of advanced econometric and machine learning techniques, based on an overall sample spanning from 18 July 2010 to 5 May 2023. The research is systematically organised into three distinct yet interconnected studies, each contributing uniquely to the understanding of market dynamics.
The first study significantly expands the empirical foundation of prior research by examining a wide range of potential return determinants across 25 leading cryptocurrencies, including well-established assets such as BTC, ETH, and XRP, which are categorised under network, production, macroeconomic, cryptocurrency-specific, and global or uncertainty factors. This approach enables a multi-dimensional analysis of cryptocurrency price formation and supports the evaluation of several theoretical perspectives, including supply and demand, cost of production, and network effects, within the context of digital assets. Methodologically, the study employs individual-level analyses, a market index-based approach, and a panel setting to capture aggregate dynamics. Principal Component Analysis is used to reduce noise among correlated variables and to identify the most salient factors across the sample. Estimations are conducted using a variety of econometric techniques, such as OLS, regression with Newey–West standard errors and autocorrelation correction, Fixed Effects, FGLS, and Prais–Winsten regression with panel-corrected standard errors.
The findings demonstrate that, although the relative importance of each factor varies across different cryptocurrencies, network-related and cryptocurrency-specific variables consistently exert the strongest influence. From a theoretical perspective, while each framework based on the analysed set of factors provides valuable insights, the results indicate that no single set of variables or theoretical lens is sufficient to fully explain cryptocurrency returns. Drawing on these results, the research establishes a novel link between the subjective theory of value and cryptocurrencies, suggesting that the value of digital assets is not derived from intrinsic fundamentals but is instead shaped by factors such as user interest, transactional benefits, and the willingness to adopt, which in turn influence network size, trading volume, and transactional activity. This connection enables the integration of multiple theoretical perspectives, incorporating behavioural and network-based influences, and contributes to the development of a more comprehensive valuation framework tailored to the unique characteristics of digital assets. In addition, the study highlights the important role of macroeconomic factors and global uncertainty, particularly during crisis periods such as the COVID-19 pandemic. This finding challenges earlier research that deemed macroeconomic variables largely insignificant and offers empirical support for their time-varying influence. By demonstrating that the sensitivity of cryptocurrency valuations to macroeconomic and financial indicators can shift over time, the study enriches the debate on whether digital assets function purely as speculative instruments or as emerging macro-sensitive assets.
The second study extends the research by investigating the predictive capacity of the identified return determinants through out-of-sample forecasting, implementing both conventional econometric models (ARIMA/ARIMAX) and modern neural networks (NAR/NARX). Based on economic and statistical loss functions, this comparative analysis demonstrates that both methodologies can effectively forecast cryptocurrency returns over short horizons, with the highest accuracy observed for one-day-ahead predictions. Notably, neural networks outperform ARIMA models when relying solely on historical return data, while ARIMAX models exhibit strong performance and often exceed that of NARX when relevant external predictors are employed.
However, the ARIMAX model’s dependence on accurate future values of exogenous inputs presents a practical limitation, as such data are often difficult to obtain in real-world or ex-ante forecasting scenarios. In contrast, the NARX model, due to its internal feedback mechanisms, has the potential to maintain forecasting accuracy even when external inputs are unavailable, delayed, or difficult to predict. This comparison highlights the respective strengths and limitations of each modelling approach and offers practical guidance for investors, traders, and analysts in selecting appropriate forecasting methods based on data availability and forecasting objectives. Moreover, the study contributes to the literature by demonstrating the feasibility of predictive modelling in cryptocurrency markets, thereby challenging the weak and semi-strong forms of market efficiency. It also reinforces the predictive relevance of network-related and cryptocurrency-specific variables, emphasising their central role in return dynamics.
The third study makes a significant contribution to the cryptocurrency literature by examining how investor heterogeneity shapes volatility, highlighting the distinct roles of institutional and individual behaviour. It employs traditional GARCH-family models alongside advanced neural networks and support vector regression techniques using diverse model architectures. The study finds that institutional investor attention consistently increases volatility, driven by access to capital, superior information, and advanced analytical tools, positioning institutions as dominant yet destabilising participants. This challenges the assumption that institutional presence enhances market stability. In contrast, individual investor attention tends to amplify volatility during stable periods but reduces it during times of global uncertainty, revealing context-dependent behavioural shifts.
The study also presents a novel multi-asset investigation into the "whale" phenomenon, demonstrating the potential of whales and large institutions to exploit informational asymmetries and amplify market inefficiencies, contributing to various forms of herding behaviours. By uniquely framing these dynamics within the context of imperfect market competition, the findings challenge the commonly held notion of decentralisation in cryptocurrency markets. This perspective highlights the risks posed by concentrated market power and the potential for manipulation, reinforcing the importance of transparency and regulatory oversight.
Moreover, the observed interdependence among cryptocurrencies, particularly the influence of Bitcoin-related institutional activity and whale behaviour on the broader market index, enhances understanding of systemic risk by showing how dominant asset-specific factors can shape overall market dynamics. Finally, the study advances volatility modelling by demonstrating that machine learning methods outperform traditional GARCH models in predictive accuracy, measured by RMSE and MAE, with neural networks proving the most effective. Their ability to capture complex, non-linear, dynamic, and high-dimensional relationships highlights the limitations of conventional models and underscores the value of data-driven techniques for modelling the unique characteristics of digital assets.
Taken together, this thesis substantially enhances the current literature by integrating theoretical perspectives with empirical findings that illuminate the distinct market mechanisms inherent in digital assets. Theoretically, it advances the understanding of how subjectivist valuation, behavioural finance, and investor heterogeneity shape the cryptocurrency market. Practically, for investors and traders, the refined predictive models and the highlighted significance of network and cryptocurrency-specific factors, attention, and whale metrics offer practical insights for portfolio optimisation and algorithmic trading strategies. For policymakers and regulators, the insights into market inefficiencies, the impact of investor attention, and the destabilising potential of institutional trades and whale activity provide critical evidence to inform regulatory reforms aimed at controlling concentration levels and stabilising this volatile market. Furthermore, the results may prove valuable for central banks exploring the design and implementation of digital currencies, as they highlight the multifaceted drivers of cryptocurrency returns and volatility and the practical considerations involved in regulating emerging financial technologies.
| Item Type: |
Thesis (University of Nottingham only)
(PhD)
|
| Supervisors: |
Haile, Getinet
Chen, Minjia |
| Keywords: |
Cryptocurrencies; Returns; Volatility; Investor Attention; Machine Learning |
| Subjects: |
H Social sciences > HG Finance |
| Faculties/Schools: |
UK Campuses > Faculty of Social Sciences, Law and Education > Nottingham University Business School |
| Item ID: |
81074 |
| Depositing User: |
Gadirli, Farid
|
| Date Deposited: |
09 Jun 2025 11:00 |
| Last Modified: |
09 Jun 2025 11:00 |
| URI: |
https://eprints.nottingham.ac.uk/id/eprint/81074 |
Actions (Archive Staff Only)
 |
Edit View |
Tools
Tools
![[thumbnail of Submission with minor corrections]](https://eprints.nottingham.ac.uk/style/images/fileicons/application_pdf.png "Submission with minor corrections")