In quantitative finance, stock trading methodologies are concerned with the definition of proper investment actions in view of maximizing a specified investor’s utility function over a defined time horizon. Most of the existing approaches in financial practice aim at deriving forecast models to accurately predict future price returns, in order to adequately take effective investment decisions. However, due to the intrinsic high non-stationarity characterizing asset price dynamics, reliable future predictions are hardly achievable even resorting to complex modeling tools and negligible prediction errors may cause a huge detrimental impact on investor’s utility. Alternatively to the mentioned traditional model-based strategies, novel trading policies relying on different return dynamics formulations have been recently theorized in control systems community. In this thesis, innovative control theoretical trading approaches are proposed, with the objective of deriving practical solutions to several problems of interest in modern quantitative finance. The common thread characterizing the presented methods concerns the exploitation of data-driven adaptive control techniques intended to overcome the literary issues connected to the prediction of price returns, by introducing trading methods independent on critical prior return model assumptions and resorting to a breakthrough trading-oriented identification rationale. The dissertation initially focuses on single-stock trading framework and provides a novel adaptive strategy inspired by reactive trading theory and literary Simultaneous Long-Short (SLS) control architecture introduced by Prof. B. R. Barmish. The underlying principle consists in re-formulating stock trading as a model-free feedback loop control problem, where stock price dynamics represents an unknown disturbance to be rejected. Despite the rigorous mathematical formulation and the proven theoretical properties when specific canonical models are assumed for price returns, literary SLS architecture presents some limitations from a practical perspective. Specifically, static controller tuning provided for the original formulation has been shown to be not feasible for real-market applications. To overcome such criticality, in this work a new data-driven adaptive methodology to iteratively tune SLS controllers is proposed, exploiting the available information about price and gain/loss realizations over the most recent past moving window, assumed as illustrative of the current market evolution. The resulting strategy is completely independent on any return model assumption and, accordingly, perfectly suitable for real-world trading. The proposed adaptive paradigm is, then, generalized to pairs trading framework, representing a further subject of notable interest in financial practice. The idea is to adopt a Cross-Coupled SLS (CC-SLS) architecture to exploit possible correlations between potentially cross-coupled stocks, usually selected from the same reference market, with the purpose of maximizing the overall investor’s utility. The major innovation compared to existing literature relates to a dynamic procedure finalized to iteratively estimate cross-coupling coefficient between the considered stocks and to adaptively optimize CC-SLS control parameter set, thus guaranteeing a full practical feasibility in real-market scenarios. Finally, the most general problem of optimal allocation in multi-asset portfolios is addressed. In detail, a Learning Model Predictive Control (MPC) scheme is formulated to solve multi-period portfolio optimization based on receding horizon principle. The main novelty regards the trading-oriented paradigm adopted to identify return prediction model, inspired by Identification for Control principle. Assuming a pre-defined model structure for returns, the idea is to identify model parameters that maximize investor’s utility, representing the actual output of interest in MPC scheme, instead of the set of parameters minimizing prediction error as in traditional Ordinary Least Squares methods. A back-testing analysis conducted on a real-world multi-asset portfolio shows the potential improvements introduced by such innovative model identification paradigm. The derived trading methodologies are extensively validated using real-market historical quotations, revealing significant outperformance in terms of investment results compared to the state of the art and some benchmark approaches from financial practice.
In finanza quantitativa, le metodologie di stock trading riguardano la definizione di opportune azioni di investimento al fine di massimizzare la funzione di utilità di uno specifico investitore su un orizzonte temporale definito. La maggior parte degli approcci esistenti nella pratica finanziaria mira a derivare modelli previsionali per prevedere in modo accurato i rendimenti futuri dei prezzi, al fine di prendere decisioni di investimento efficaci. Tuttavia, a causa dell'elevata non-stazionarietà intrinseca che caratterizza la dinamica dei prezzi degli asset, previsioni affidabili sono difficilmente realizzabili anche ricorrendo a strumenti di modellazione complessi ed errori di predizione trascurabili possono causare un enorme impatto negativo sull'utilità dell'investitore. In alternativa alle strategie tradizionali model-based, nella comunità dei sistemi di controllo sono state recentemente teorizzate nuove politiche di trading. In questa tesi vengono proposti approcci di trading innovativi, con l'obiettivo di derivare soluzioni pratiche a diversi problemi di interesse nella finanza quantitativa moderna. Il filo conduttore che caratterizza i metodi presentati riguarda l’utilizzo di tecniche di controllo adattativo volte a superare le problematiche di letteratura legate alla predizione dei rendimenti dei prezzi, introducendo metodi di trading indipendenti dalle ipotesi critiche e ricorrendo ad un nuovo paradigma di identificazione. Il lavoro si concentra inizialmente sul framework di single-stock trading e propone una nuova strategia adattativa ispirata alla teoria del reactive trading e all'architettura di controllo Simultaneous Long-Short (SLS) introdotta dal Prof. B. R. Barmish. Il principio alla base consiste nel riformulare lo stock trading come un problema di controllo model-free, in cui la dinamica del prezzo delle azioni rappresenta un disturbo. Nonostante la rigorosa formulazione matematica e le comprovate proprietà teoriche nell’ipotesi di modelli canonici specifici per i rendimenti dei prezzi, l'architettura SLS tradizionale presenta alcune limitazioni in una prospettiva pratica. In particolare, la taratura del controllore statico prevista nella formulazione originale non è adeguata per applicazioni del mercato reale. Al fine di ovviare a tale criticità, in questa tesi viene proposta una nuova metodologia adattativa data-driven per la taratura iterativa dei controllori SLS, sfruttando le informazioni disponibili relative alle realizzazioni di prezzi e guadagni/perdite sulla più recente finestra mobile passata, assunta come illustrativa dell'attuale evoluzione del mercato. La strategia risultante è completamente indipendente da qualsiasi ipotesi modellistica e, di conseguenza, perfettamente adatta per il real-market trading. Il paradigma adattativo proposto è, quindi, generalizzato al framework del pairs trading, che rappresenta un ulteriore argomento di notevole interesse nella pratica finanziaria. L'idea è quella di adottare un'architettura Cross-Coupled SLS (CC-SLS) per sfruttare eventuali correlazioni tra titoli potenzialmente cross-coupled, solitamente selezionati dallo stesso mercato di riferimento, con l'obiettivo di massimizzare l'utilità complessiva dell'investitore. La principale innovazione rispetto allo stato dell’arte riguarda una procedura dinamica finalizzata a stimare iterativamente il coefficiente di cross-coupling tra i titoli considerati e ad ottimizzare in modo adattativo il set di parametri di controllo CC-SLS, garantendo così una piena applicabilità in scenari di mercato reale. Infine, viene affrontato il problema più generale dell'allocazione ottima in portafogli multi-asset. Nello specifico, viene formulato uno schema Model Predictive Control (MPC) per risolvere il problema di portfolio optimization multi-periodo basata sul principio del receding horizon. La principale innovazione riguarda il paradigma trading-oriented adottato per identificare il modello di predizione dei rendimenti, ispirato al principio di Identification for Control. Assumendo una struttura del modello predefinita per i rendimenti, l'idea è identificare i parametri del modello che massimizzano l'utilità dell'investitore, che costituisce l’effettivo output di interesse nello schema MPC, anziché il set di parametri che minimizzano l'errore di predizione come nei metodi tradizionali Ordinary Least Squares. Un'analisi di back-testing condotta su un portafoglio multi-asset reale mostra i potenziali miglioramenti introdotti da tale paradigma innovativo per l’identificazione modellistica. Le metodologie di trading presentate sono validate utilizzando quotazioni storiche del mercato reale, mostrando un significativo miglioramento delle performance in termini di risultati di investimento rispetto allo stato dell'arte ed alcuni approcci benchmark dalla pratica finanziaria.
Reactive trading : a control theoretical approach to financial engineering
ABBRACCIAVENTO, FRANCESCO
2021/2022
Abstract
In quantitative finance, stock trading methodologies are concerned with the definition of proper investment actions in view of maximizing a specified investor’s utility function over a defined time horizon. Most of the existing approaches in financial practice aim at deriving forecast models to accurately predict future price returns, in order to adequately take effective investment decisions. However, due to the intrinsic high non-stationarity characterizing asset price dynamics, reliable future predictions are hardly achievable even resorting to complex modeling tools and negligible prediction errors may cause a huge detrimental impact on investor’s utility. Alternatively to the mentioned traditional model-based strategies, novel trading policies relying on different return dynamics formulations have been recently theorized in control systems community. In this thesis, innovative control theoretical trading approaches are proposed, with the objective of deriving practical solutions to several problems of interest in modern quantitative finance. The common thread characterizing the presented methods concerns the exploitation of data-driven adaptive control techniques intended to overcome the literary issues connected to the prediction of price returns, by introducing trading methods independent on critical prior return model assumptions and resorting to a breakthrough trading-oriented identification rationale. The dissertation initially focuses on single-stock trading framework and provides a novel adaptive strategy inspired by reactive trading theory and literary Simultaneous Long-Short (SLS) control architecture introduced by Prof. B. R. Barmish. The underlying principle consists in re-formulating stock trading as a model-free feedback loop control problem, where stock price dynamics represents an unknown disturbance to be rejected. Despite the rigorous mathematical formulation and the proven theoretical properties when specific canonical models are assumed for price returns, literary SLS architecture presents some limitations from a practical perspective. Specifically, static controller tuning provided for the original formulation has been shown to be not feasible for real-market applications. To overcome such criticality, in this work a new data-driven adaptive methodology to iteratively tune SLS controllers is proposed, exploiting the available information about price and gain/loss realizations over the most recent past moving window, assumed as illustrative of the current market evolution. The resulting strategy is completely independent on any return model assumption and, accordingly, perfectly suitable for real-world trading. The proposed adaptive paradigm is, then, generalized to pairs trading framework, representing a further subject of notable interest in financial practice. The idea is to adopt a Cross-Coupled SLS (CC-SLS) architecture to exploit possible correlations between potentially cross-coupled stocks, usually selected from the same reference market, with the purpose of maximizing the overall investor’s utility. The major innovation compared to existing literature relates to a dynamic procedure finalized to iteratively estimate cross-coupling coefficient between the considered stocks and to adaptively optimize CC-SLS control parameter set, thus guaranteeing a full practical feasibility in real-market scenarios. Finally, the most general problem of optimal allocation in multi-asset portfolios is addressed. In detail, a Learning Model Predictive Control (MPC) scheme is formulated to solve multi-period portfolio optimization based on receding horizon principle. The main novelty regards the trading-oriented paradigm adopted to identify return prediction model, inspired by Identification for Control principle. Assuming a pre-defined model structure for returns, the idea is to identify model parameters that maximize investor’s utility, representing the actual output of interest in MPC scheme, instead of the set of parameters minimizing prediction error as in traditional Ordinary Least Squares methods. A back-testing analysis conducted on a real-world multi-asset portfolio shows the potential improvements introduced by such innovative model identification paradigm. The derived trading methodologies are extensively validated using real-market historical quotations, revealing significant outperformance in terms of investment results compared to the state of the art and some benchmark approaches from financial practice.File | Dimensione | Formato | |
---|---|---|---|
PhD Thesis - Francesco Abbracciavento.pdf
non accessibile
Descrizione: Ph.D. Thesis - Francesco Abbracciavento
Dimensione
2.12 MB
Formato
Adobe PDF
|
2.12 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/10589/187542