This thesis investigates whether Long Short-Term Memory (LSTM) networks can improve the portfolio optimization by providing forward-looking estimates of stock returns within a traditional Markowitz mean–variance framework. The analysis is conducted on 45 stocks taken from the EURO STOXX 50, by using daily adjusted prices and trading volumes from January 2015 to May 2026. The LSTM model was trained to predict three-day cumulative log-returns using price, momentum, volume, and volatility based features. The resulting forecasts are then used to construct a complete and a filtered LSTM portfolios, which are compared with a historical Markowitz portfolio, an equally weighted portfolio, and the EURO STOXX 50 benchmark. Although the LSTM shows only modest standalone forecasting accuracy, the portfolio results indicate that its predictions contain economically useful information. In the out-of-sample back-test, the filtered LSTM portfolio achieves the strongest performance, with an annualized return of 66.8% and a Sharpe ratio of 3.14 after transaction costs, outperforming all alternative strategies. Overall, the findings suggest that LSTM forecasts can enhance portfolio allocation when combined with appropriate stock-selection criteria and traditional risk estimation.
Questa tesi analizza se le reti Long Short-Term Memory (LSTM) possano migliorare l’ottimizzazione di portafoglio fornendo stime prospettiche dei rendimenti azionari all’interno del tradizionale framework media-varianza di Markowitz. L’analisi è condotta su 45 titoli appartenenti all’EURO STOXX 50, utilizzando prezzi di chiusura rettificati giornalieri e volumi di negoziazione da gennaio 2015 a maggio 2026. Il modello LSTM è stato addestrato per prevedere i rendimenti logaritmici cumulati su tre giorni utilizzando variabili basate su prezzo, momentum, volume e volatilità. Le previsioni ottenute vengono successivamente utilizzate per costruire un portafoglio LSTM completo e un portafoglio LSTM filtrato, confrontati con un portafoglio storico di Markowitz, un portafoglio equamente ponderato e con l’indice EURO STOXX 50 utilizzato come benchmark. Sebbene il modello LSTM mostri una capacità predittiva individuale piuttosto modesta, i risultati a livello di portafoglio indicano che le sue previsioni contengono informazioni economicamente utili. Nel back-test out-of-sample, il portafoglio LSTM filtrato registra la performance migliore, con un rendimento annualizzato del 66,8% e uno Sharpe ratio pari a 3,14 dopo l’inclusione dei costi di transazione, superando tutte le strategie alternative. Nel complesso, i risultati suggeriscono che le previsioni LSTM possono migliorare l’allocazione di portafoglio quando vengono combinate con opportuni criteri di selezione dei titoli e con una metodologia tradizionale di stima del rischio.
Portfolio optimization with a Machine Learning model: The LSTM Network
PARISI, MARCO
2025/2026
Abstract
This thesis investigates whether Long Short-Term Memory (LSTM) networks can improve the portfolio optimization by providing forward-looking estimates of stock returns within a traditional Markowitz mean–variance framework. The analysis is conducted on 45 stocks taken from the EURO STOXX 50, by using daily adjusted prices and trading volumes from January 2015 to May 2026. The LSTM model was trained to predict three-day cumulative log-returns using price, momentum, volume, and volatility based features. The resulting forecasts are then used to construct a complete and a filtered LSTM portfolios, which are compared with a historical Markowitz portfolio, an equally weighted portfolio, and the EURO STOXX 50 benchmark. Although the LSTM shows only modest standalone forecasting accuracy, the portfolio results indicate that its predictions contain economically useful information. In the out-of-sample back-test, the filtered LSTM portfolio achieves the strongest performance, with an annualized return of 66.8% and a Sharpe ratio of 3.14 after transaction costs, outperforming all alternative strategies. Overall, the findings suggest that LSTM forecasts can enhance portfolio allocation when combined with appropriate stock-selection criteria and traditional risk estimation.| File | Dimensione | Formato | |
|---|---|---|---|
|
PARISI_MARCO_0905951.pdf
accesso aperto
Dimensione
1.19 MB
Formato
Adobe PDF
|
1.19 MB | Adobe PDF | Visualizza/Apri |
I documenti in UNITESI sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/20.500.14247/30241