Essays on learning and memory in virtual currency markets

Abstract

Learning is vital in a growing system, irrespective of disciplinary distinctions and discrimination between physical and non-physical spaces. A central force that guides the process and speed of learning is the magnitude of memory that the system inherits and one that purportedly interacts with external factors. Although in the broad field of finance and economics, the vitality of learning and memory is now an established mechanic that regularly elicits empirical reliability of inferences, its efficacy in the virtual currency market is still a contested domain. Worse still, it appears to be lost in the sinuous world of transitional arguments of the underlying foundation of financial theory in virtual currencies. As such, there is no fundamental value attached to a virtual currency market and, therefore, what must drive the predictive power in this market and persistent positive return is to have a robust model of learning and memory. Among the constant features of randomness as the driver of virtual currency markets, what is more certain here is the way agents learn and model memory. This thesis aims to contribute to the nascent literature by conducting an in-depth empirical and theoretical review of the role of learning and memory in virtual currency markets. Lately, a substantial and growing body of research has focused on long memory in cryptocurrency markets, highlighting market inefficiency. Furthermore, recent advances in machine learning (ML) algorithms have demonstrated their predictive accuracy, suggesting that markets may be exploitable. Deep-reinforcement learning (DRL) frameworks also offer promising potential for developing effective trading strategies. However, there is a notable gap in the literature on exploration of univariate and systemic memory dynamics in the context of multiple frequencies, event shocks, and structural breaks. Moreover, there is a lack of approaches to evaluate ML models’ ability to process long-memory series and a need for a comprehensive exploration of how DRL can improve strategy performance. This thesis addresses these gaps by analysing the timevarying properties of memory, designing experiments to model long-memory series to evaluate prevalent ML models, and investigating how memory indicators and delayed rewards can influence strategy performance. These contributions offer valuable insights into trading within inefficient markets and establish an analytical framework for iv the cryptocurrency market that integrates Finance, Statistics, and Artificial Intelligence (AI). Chapter 2 proposes a comprehensive analysis framework for evaluating univariate and systemic long memory in financial markets. Basically, it focusses on market efficiency and employs novel methods to examine the memory properties of Bitcoin (BTC) and Ethereum (ETH) prices on different frequencies in cryptocurrency markets. Specifically, we aggregated high-frequency transaction data, comprising 1,188,000 observations obtained from Binance, into hourly, daily, and weekly frequencies. Using the feasible exact local whittle (FELW) estimate, we assess the long-memory properties or persistence of BTC and ETH prices. Furthermore, we introduce an innovative approach to explore the time-varying systemic memory within the microstructure of price changes. This is achieved by integrating the rolling window technique with the Fractionally Cointegrated Vector Autoregressive (FCVAR) model. The empirical results reveal the presence of time-varying market efficiency and dynamic predictability in the two leading cryptocurrencies, BTC and ETH. These findings suggest that inefficiency is a natural state in cryptocurrency markets, despite careful consideration of structural breaks and regime-switching effects on memory properties. This analysis not only contributes to the understanding of long-memory dynamics in cryptocurrency markets, but also enhances the methodologies available for assessing market efficiency and predictability, paving the way for more robust financial modelling and forecasting techniques. Chapter 3 focusses on examining the ability of seven widely used ML regression algorithms, along with sequence-to-sequence (Seq2Seq) models, to capture long-memory characteristics in financial data. This capability is evaluated from two key perspectives. First, we analyse whether these algorithms can accurately estimate the fractional integration parameter d compared to established methods such as the FELW estimator and the FCVAR model. Second, we investigate whether the time series predicted by these models exhibit similar long-memory properties to the original data. Our results show that most of the ML algorithms evaluated fail to effectively handle long-memory series, whereas models incorporating LSTM and Attention-LSTM components demonstrate superior performance. This finding highlights the limitations of conventional ML models in dealing with financial time series characterised by long-range dependence, suggesting that Seq2Seq models may be more suitable for modelling such series. These results emphasise the challenges of applying ML techniques to financial time series data with long-memory properties and point to the need for further research and development of specialised algorithms that can better address these complexities. This research contributes to a deeper understanding of the intersection between ML and econometrics, shedding light on the areas where traditional econometric methods and modern ML approaches may differ in their effectiveness when applied to financial data. Chapter 4 focusses on developing and optimising trend trading strategies within the framework of multiple frequencies, long-memory effects, and delayed rewards. The v chapter investigates the ideal market conditions for the use of momentum indicators and optimises these strategies by incorporating memory indicators, ML, and DRL techniques. The findings indicate that using a sampling interval of 40,320 minutes to calculate momentum can lead to significantly higher profitability. Strategies optimised through ML methods, particularly with a three-layer neural network (NN) model, demonstrate robust performance, achieving a 300% return on the test data set. Although memory indicators show potential to improve strategy performance, their effectiveness is challenging to exploit as only specific combinations of factors result in improved results, while others may detract from performance. This suggests that the interaction between memory and momentum indicators is highly context-dependent. Moreover, both the ML and DRL methods struggle to consistently identify these optimal combinations, indicating that the relationship may be time-varying and difficult to capitalise on systematically. Furthermore, the results reveal that simple neural network models can outperform more complex algorithms, such as the Proximal Policy Optimisation (PPO) model, which is often considered a sophisticated DRL approach. Surprisingly, the PPO models incorporating LSTM-Attention mechanisms underperform compared to even the benchmark models. This suggests that, in the context of trend trading in cryptocurrency markets, simpler models may offer greater robustness and reliability than more advanced architectures that are prone to overfitting or capturing noise. In general, the chapter highlights the complexities and challenges involved in optimising trend trading strategies with long-memory and ML techniques. It underscores the need for further research to better understand how to effectively integrate memory indicators and ML models to exploit market inefficiencies that vary over time. These findings suggest that, while ML and DRL offer powerful tools for financial modelling, their application in markets with dynamic characteristics, such as cryptocurrencies, requires careful consideration of model simplicity, robustness, and adaptability.

More information

Identifiers

ORCID for Shuyue Li: orcid.org/0009-0002-6671-7162

ORCID for Tapas Mishra: orcid.org/0000-0002-6902-2326

ORCID for Larisa Yarovaya: orcid.org/0000-0002-9638-2917

Catalogue record

Export record