The aim of this thesis project consists in developing a trading system on cryptocurrencies, through the use of artificial intelligence, incorporating sentiment analysis of Twitter micro-blogging contents and other financial features. To achieve this task the development of the project evolved in three phases. The first one consists in finding a way to calculate the sentiment expressed by the millions of Tweets posted along the years related to cryptocurrencies' topics. This issue was solved using a parsing algorithm, able to download and transcribe the necessary tweets, and the VADER dictionary, to determine the sentiment expressed. The second phase consists in analysing whether exogenous features, such as those related to sentiment and other financial features, might have some correlations with the considered cryptocurrencies historical price and in investigating the established relationships among different cryptocurrencies. It emerged that Bitcoin can still be considered as a benchmark for the total cryptocurrency market, even if different relationships are beginning to emerge with the increasing adoption and development of this sector. Furthermore, the extraordinary macroeconomic conditions after Covid-19 outbreak created some distortions in the market and investors' sentiment, increasing the overall correlation between different financial assets, especially between Bitcoin and the major US stock indexes. In the third phase different trading systems using artificial intelligence were developed and tested. After the initial failed attempt, in which basic machine learning models were implemented reporting performances well below expectations, they were substituted by recurrent neural networks, more akin to the objectives set. The ultimate goal of this project is the development of a realistic trading system capable of beating the Buy & Hold strategy benchmark performed on multiple cryptocurrencies such as Bitcoin, Ethereum, Cardano and Polkadot. Furthermore, it has been verified that the inclusion of exogenous features in the datasets, in addition to the mere price action data, is able to provide meaningful information to the AI models, increasing their predictive capacity.
L'obiettivo di questo progetto di tesi consiste nello sviluppare un sistema di trading sulle criptovalute, attraverso l'uso dell'intelligenza artificiale, che incorpori l'analisi del sentiment dei contenuti di micro-blogging di Twitter ed altre variabili finanziarie. Per raggiungere questo scopo lo sviluppo del progetto si è evoluto in tre fasi. La prima consiste nel trovare un modo per calcolare il sentiment espresso dai milioni di Tweet pubblicati negli anni relativi ad argomenti sulle criptovalute. Le problematiche relative a questa fase sono state risolte utilizzando un algoritmo di parsing, in grado di scaricare e trascrivere i tweet necessari, e il dizionario VADER, per determinare il sentiment espresso. La seconda fase consiste nell'analizzare se le variabili esogene, come quelle legate al sentiment e ad altre variabili finanziarie, riportassero delle correlazioni con il prezzo storico delle criptovalute considerate e nell'investigare le relazioni che si sono stabilite tra le differenti criptovalute. È emerso che Bitcoin può ancora essere considerato il punto di riferimento e fulcro per tutto il mercato delle criptovalute, anche se con la crescente adozione e sviluppo di questo settore, differenti relazioni iniziano ad emergere e diventare evidenti. Inoltre, le straordinarie condizioni macroeconomiche successive all'epidemia di Covid-19 hanno creato alcune distorsioni nel mercato e nel sentimento degli investitori, aumentando la correlazione complessiva tra i diversi asset finanziari, in particolare tra Bitcoin e i principali indici azionari statunitensi. Nella terza fase sono stati sviluppati e testati diversi sistemi di trading che utilizzano l'intelligenza artificiale. Dopo il fallito tentativo iniziale, in cui sono stati implementati modelli ML base che hanno riportato prestazioni ben al di sotto delle aspettative, essi sono stati sostituiti da reti neurali ricorrenti, più affini agli scopi del progetto. L'obiettivo finale di questo progetto è lo sviluppo di un sistema di trading realistico in grado di battere il benchmark della strategia Buy&Hold eseguita su più criptovalute, come Bitcoin, Ethereum, Cardano e Polkadot. Inoltre, è stato verificato che l'inclusione di variabili esogene nei dataset, oltre ai meri dati relativi all'andamento dei prezzi e dei volumi di scambio, è in grado di fornire informazioni significative ai modelli di AI, aumentandone la capacità predittiva.
Development of a trading system on cryptocurrencies using machine learning exploiting sentiment analysis
Beatrice, Antonio;NOBILI, FEDERICO
2020/2021
Abstract
The aim of this thesis project consists in developing a trading system on cryptocurrencies, through the use of artificial intelligence, incorporating sentiment analysis of Twitter micro-blogging contents and other financial features. To achieve this task the development of the project evolved in three phases. The first one consists in finding a way to calculate the sentiment expressed by the millions of Tweets posted along the years related to cryptocurrencies' topics. This issue was solved using a parsing algorithm, able to download and transcribe the necessary tweets, and the VADER dictionary, to determine the sentiment expressed. The second phase consists in analysing whether exogenous features, such as those related to sentiment and other financial features, might have some correlations with the considered cryptocurrencies historical price and in investigating the established relationships among different cryptocurrencies. It emerged that Bitcoin can still be considered as a benchmark for the total cryptocurrency market, even if different relationships are beginning to emerge with the increasing adoption and development of this sector. Furthermore, the extraordinary macroeconomic conditions after Covid-19 outbreak created some distortions in the market and investors' sentiment, increasing the overall correlation between different financial assets, especially between Bitcoin and the major US stock indexes. In the third phase different trading systems using artificial intelligence were developed and tested. After the initial failed attempt, in which basic machine learning models were implemented reporting performances well below expectations, they were substituted by recurrent neural networks, more akin to the objectives set. The ultimate goal of this project is the development of a realistic trading system capable of beating the Buy & Hold strategy benchmark performed on multiple cryptocurrencies such as Bitcoin, Ethereum, Cardano and Polkadot. Furthermore, it has been verified that the inclusion of exogenous features in the datasets, in addition to the mere price action data, is able to provide meaningful information to the AI models, increasing their predictive capacity.| File | Dimensione | Formato | |
|---|---|---|---|
|
Tesi_Beatrice_Nobili.pdf
accessibile in internet solo dagli utenti autorizzati
Dimensione
12.86 MB
Formato
Adobe PDF
|
12.86 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/185824