In recent years, Streaming Machine Learning (SML) has gained increasing attention due to its ability to process data incrementally and adapt models in real time. However, its application to financial time series forecasting remains less explored compared to traditional batch-based Deep Learning approaches. Financial markets are inherently non-stationary and characterized by volatility and distributional shifts, which can significantly affect predictive performance. This thesis presents a comparative study between a state-of-the-art Deep Learning model, the Time-series Dense Encoder (TiDE), and adaptive Streaming Machine Learning methods for short-term stock market prediction. The objective is to evaluate whether streaming approaches can provide competitive predictive accuracy while offering advantages in adaptability and computational efficiency. Experiments are conducted on the S&P500 index and the Brazilian ETF EWZ, considering both daily and hourly data. The results show that Deep Learning achieves slightly higher aggregate accuracy. Nevertheless, the streaming models consistently perform above the random baseline and demonstrate structural advantages, including real-time prediction capability and incremental adaptation to evolving data without full retraining. These findings highlight Streaming Machine Learning as a viable and efficient alternative for financial forecasting in dynamic environments.
Negli ultimi anni, lo Streaming Machine Learning (SML) ha attirato crescente attenzione grazie alla sua capacità di elaborare dati in modo incrementale e di adattare i modelli in tempo reale. Tuttavia, la sua applicazione alla previsione di serie temporali finanziarie rimane meno esplorata rispetto ai tradizionali approcci di Deep Learning basati su addestramento batch. I mercati finanziari sono intrinsecamente non stazionari e caratterizzati da volatilità e variazioni nella distribuzione dei dati, che possono influenzare significativamente le prestazioni predittive. Questa tesi presenta uno studio comparativo tra un modello di Deep Learning all’avanguardia, il Time-series Dense Encoder (TiDE), e metodi adattivi di Streaming Machine Learning per la previsione a breve termine dei prezzi azionari. L’obiettivo è valutare se gli approcci di tipo streaming possano fornire un’accuratezza predittiva competitiva offrendo al contempo vantaggi in termini di adattabilità ed efficienza computazionale. Gli esperimenti sono condotti sull’indice SP500 e sull’ETF brasiliano EWZ, considerando sia dati giornalieri sia orari. I risultati mostrano che il Deep Learning raggiunge un’accuratezza aggregata leggermente superiore. Tuttavia, i modelli di streaming mantengono prestazioni costantemente superiori alla baseline casuale e dimostrano vantaggi strutturali, tra cui la capacità di effettuare previsioni in tempo reale e di adattarsi in modo incrementale ai dati in evoluzione senza richiedere un riaddestramento completo. Questi risultati evidenziano lo Streaming Machine Learning come un’alternativa valida ed efficiente per la previsione finanziaria in ambienti dinamici.
A comparative study of Deep Learning and Streaming Machine Learning approaches for stock market forecasting
Bucaioni, Tommaso
2024/2025
Abstract
In recent years, Streaming Machine Learning (SML) has gained increasing attention due to its ability to process data incrementally and adapt models in real time. However, its application to financial time series forecasting remains less explored compared to traditional batch-based Deep Learning approaches. Financial markets are inherently non-stationary and characterized by volatility and distributional shifts, which can significantly affect predictive performance. This thesis presents a comparative study between a state-of-the-art Deep Learning model, the Time-series Dense Encoder (TiDE), and adaptive Streaming Machine Learning methods for short-term stock market prediction. The objective is to evaluate whether streaming approaches can provide competitive predictive accuracy while offering advantages in adaptability and computational efficiency. Experiments are conducted on the S&P500 index and the Brazilian ETF EWZ, considering both daily and hourly data. The results show that Deep Learning achieves slightly higher aggregate accuracy. Nevertheless, the streaming models consistently perform above the random baseline and demonstrate structural advantages, including real-time prediction capability and incremental adaptation to evolving data without full retraining. These findings highlight Streaming Machine Learning as a viable and efficient alternative for financial forecasting in dynamic environments.| File | Dimensione | Formato | |
|---|---|---|---|
|
2026_03_Bucaioni.pdf
accessibile in internet per tutti
Descrizione: testo della tesi
Dimensione
694.08 kB
Formato
Adobe PDF
|
694.08 kB | 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/10589/251883