Early Warning System Using Machine Learning Techniques - Application to Financial Data and Cryptocurrencies - Master’s Thesis

Aug 2, 2026

Uploaded by Rowan

This master’s thesis investigates an Early Warning System for financial markets through machine learning, emphasizing the PCA NN (Principal Component Analysis and Neural Networks) framework. The study aims to classify financial price movements as normal or abnormal using historical data, supporting applications in trading and risk management. A PCA-based feature extraction step feeds a neural network that calibrates anomaly thresholds, then results are benchmarked against Random Forest, KNN, and Autoencoder on both financial and cryptocurrency datasets, showing stronger performance on cryptocurrencies.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages58

Reading Time2 h 26 min

AccessFree to download

Introduction

Early Detection Model

PCA NN Algorithm

Results obtained with benchmark models

Conclusions and future developments

Appendix

Appendix A

Appendix B

List of Figures

List of Tables

EARLY WARNING SYSTEM USING MACHINE LEARNING TECHNIQUES: APPLICATION TO FINANCIAL DATA AND CRYPTOCURRENCIES Tesi di Laurea Magistrale in Mathematical Engineering-Quantitative Finance Author: SAMUELE ANGELO PIEMONTI Student ID: 101705 Advisor: Daniele Marazzina Co-advisors: Raffaele Zenti Academic Year: 2023-2024 i Abstract This master’s thesis explores the implementation of an Early Warning System in financial markets using machine learning techniques, with a specific focus on the "Principal Component Analysis and Neural Networks (PCA NN)" model. The objective of these models is to classify financial price movements as either normal or abnormal based on historical data, a capability that proves valuable in financial contexts such as trading and risk management. Initially, the study examines a model that utilizes PCA to extract principal components, followed by a neural network to calibrate the anomaly thresholds. This approach is then compared to benchmark models such as Random Forest, K-Nearest Neighbors (KNN), and Autoencoder. These models are applied to both financial and cryptocurrency datasets. Comparative results demonstrate that PCA NN performs well across different datasets, showing performance in line with Random Forest and KNN; while Autoencoder underperforms them. Furthermore, the results highlight that the performance of the models is generally better on the cryptocurrency dataset. The findings emphasize the importance of using multiple machine learning models to improve the accuracy and reliability of market anomaly detection systems. In conclusion, it is possible to consider an ensemble model, as the combination of these models, to implement a robust Early Warning System that is effective across both financial and cryptocurrency datasets. Keywords: Early Warning System, machine learning, PCA NN, Random Forest, KNN, financial dataset, cryptocurrencies dataset. Abstract in italiano Questa tesi magistrale esplora l’implementazione di un sitema di Early Warning nei mercati finanziari usando tecniche di machine learning, con un focus specifico sul modello"Principal Component Analysis and Neural Networks (PCA NN)". L’obiettivo di questi modelli è classificare i movimenti dei prezzi finanziari come normali o anomali basandosisu dati storici, una capacità che si rivela preziosa in contesti finanziari come trading e risk management. Inizialmente, lo studio esamina un modello che utilizza la PCA per estrarre le componenti principali, seguito da una rete neurale per calibrare le soglie di anomalia. Questo approccio viene quindi confrontato con modelli di riferimento come Random Forest, K-Nearest Neighbors (KNN) e Autoencoder. Questi modelli sono applicati sia a dataset finanziariche a dataset di criptovalute. I risultati comparativi dimostrano che il modello PCA NN performa bene sui diversi dataset, mostrando prestazioni in linea con Random Forest e KNN, mentre l’Autoencoder presenta prestazioni inferiori. Inoltre, i risultati evidenziano che le prestazioni dei modelli sono generalmente migliori sui dataset di criptovalute. Le conclusioni sottolineanol’importanza di utilizzare più modelli di machine learning per migliorare l’accuratezza el’affidabilità dei sistemi di rilevamento delle anomalie di mercato. In conclusione, è possibile considerare un modello ensemble, come combinazione di questi modelli, per implementare un sistema Early Warning robusto ed efficace su entrambi idataset finanziari e di criptovalute. Parole chiave: Early Warning System, machine learning, PCA NN, Random Forest, KNN, dataset finanziario, dataset di criptovalute. v Contents Abstract i Abstract in italiano iii Contents v Introduction 1 1 Early Detection Model 5 1.1 Presentation of the considered models ..................... 5 1.1.1 PCA NN "Principal Component Analysis and Neural Network" .. 7 1.1.2 Random Forest ............................. 9 1.1.3 KNN "K-nearest neighbors classifier" .....

FAQs

What is the main goal of the thesis?

To implement an Early Warning System that detects abnormal market behavior by classifying price movements as normal or abnormal using machine learning on historical data.

How does the PCA NN model work in this study?

PCA extracts principal components from the data, and a neural network calibrates anomaly thresholds to distinguish normal from abnormal movements.

Which models are compared and what do the results show?

The PCA NN approach is compared with Random Forest, KNN, and Autoencoder on financial and cryptocurrency datasets; PCA NN is competitive with Random Forest and KNN, while Autoencoder underperforms, with generally better performance on cryptocurrency data.

Aug 2, 2026

Uploaded by Rowan