The problem of asset allocation is tackled, trying to avoid the curse of dimensionality by selecting a small basket from a universe of securities, such as to guarantee real diversification and preserve good portfolio performance. The selection criterion is based on the mimicry of the strategies that exploit the $\beta$ anomaly, however based on a more general concept of $\beta$ that is proposed. Starting from a universe of stocks and momentum characteristics, a market model is constructed on the basis of a handful of endogenous factors obtained through the non-linear structure of an autoencoder neural network. Based on these factors, the securities are reproduced using linear models and sorted by their mean squared error of their observed value. Therefore, securities with a low Generalized Beta are defined as those for which the MSE is higher. The proposed low $\beta$ concept thus includes stocks that are poorly correlated with each other and with the market as a whole, based on the nonlinear autoencoder model. The work is therefore organized as follows: section 1 proposes a literature review on asset allocation, $\beta$ anomaly and Financial Machine Learning techniques; in the second section the autoencoder neural networks are presented in general, then the peculiar version adopted and the momentum characteristic produced are disclosed, then the portfolio selection models and the covariance estimation techniques through which it is tested are presented in detail the stock picking technique; in the third section, a main empirical application and an out-of-sample backtest, as well as some other applications, are discussed and analyzed from various perspectives; the fourth section collects the conclusions and some suggestions for future research.
Viene affrontato il problema dell'asset allocation e della maledizione della dimensionalità, selezionando un piccolo paniere da un universo di titoli, tale da garantire una reale diversificazione e preservare buone performance di portafoglio. Il criterio di selezione si basa sulla mimesi delle strategie che sfruttano l'anomalia $\beta$, basandosi però su un concetto più generale di $\beta$ che viene introdotto. Partendo da un universo di titoli e caratteristiche di momentum, viene costruito un modello di mercato sulla base di una manciata di fattori endogeni ottenuti attraverso la struttura non lineare di una rete neurale autoencoder. Sulla base di questi fattori, i titoli vengono riprodotti utilizzando modelli lineari e ordinati in base al loro errore quadratico medio rispetto al loro valore osservato. Si definiscono quindi titoli a basso Beta Generalizzato quelli per i quali il MSE è maggiore. Il concetto di $\beta$ basso proposto include in questo modo titoli scarsamente correlati tra loro e con il mercato nel suo complesso, sulla base del modello di autoencoder non lineare. Il lavoro è quindi organizzato come segue: la sezione 1 propone una revisione della letteratura su asset allocation, anomalia $\beta$ e tecniche di Financial Machine Learning; nella seconda sezione vengono presentate in generale le reti neurali di autoencoder, viene presentata la versione peculiare quivi adottata e le caratteristiche di momentum prodotte, quindi vengono presentati in dettaglio i modelli di selezione del portafoglio e le tecniche di stima della covarianza attraverso le quali viene testata la tecnica di stock picking; nella terza sezione vengono discusse e analizzate da varie prospettive diverse applicazioni empiriche e backtest, oltre ad alcune altre applicazioni; la quarta sezione raccoglie le conclusioni e alcuni suggerimenti per ricerche future.
Machine Learning: dal riconoscimento immagini ai mercati finanziari.
QYRANA, MISHEL
2023
Abstract
The problem of asset allocation is tackled, trying to avoid the curse of dimensionality by selecting a small basket from a universe of securities, such as to guarantee real diversification and preserve good portfolio performance. The selection criterion is based on the mimicry of the strategies that exploit the $\beta$ anomaly, however based on a more general concept of $\beta$ that is proposed. Starting from a universe of stocks and momentum characteristics, a market model is constructed on the basis of a handful of endogenous factors obtained through the non-linear structure of an autoencoder neural network. Based on these factors, the securities are reproduced using linear models and sorted by their mean squared error of their observed value. Therefore, securities with a low Generalized Beta are defined as those for which the MSE is higher. The proposed low $\beta$ concept thus includes stocks that are poorly correlated with each other and with the market as a whole, based on the nonlinear autoencoder model. The work is therefore organized as follows: section 1 proposes a literature review on asset allocation, $\beta$ anomaly and Financial Machine Learning techniques; in the second section the autoencoder neural networks are presented in general, then the peculiar version adopted and the momentum characteristic produced are disclosed, then the portfolio selection models and the covariance estimation techniques through which it is tested are presented in detail the stock picking technique; in the third section, a main empirical application and an out-of-sample backtest, as well as some other applications, are discussed and analyzed from various perspectives; the fourth section collects the conclusions and some suggestions for future research.File | Dimensione | Formato | |
---|---|---|---|
Mishel Qyrana - Machine Learning, from Image Recognition to Financial Markets.pdf
Open Access dal 02/09/2024
Dimensione
14.48 MB
Formato
Adobe PDF
|
14.48 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/20.500.14242/84640
URN:NBN:IT:UNIPV-84640