Francesco Abbracciavento

Major research topic

Reactive Trading: A control theoretical approach to financial engineering

Abstract

In finance, stock trading problem refers to the investigation of strategies that support investors in taking investment decisions according to the underlying price dynamics. The final aim consists in optimizing an a-priori specified gain-loss function over a reference period. Most of the existing techniques in classical financial framework intend to precisely model stock price dynamics, based on back-testing statistical analysis, for the accurate prediction of future price evolution. However, such dynamics is usually highly non-stationary and uncertain, thus even slight modelling errors may cause a huge detrimental impact on investment policies. This criticality motivated research towards alternative “model-free” strategies. In my research work, I firstly investigated novel methodologies that involve trading a single stock and then generalized to the problem of optimal share allocation in multi-asset portfolios.
Concerning single-stock trading, my research was inspired by Reactive Trading theory, that reformulates gain-loss maximization as a feedback control design problem with disturbance rejection goal. The main drawback of literary Simultaneous Long-Short (SLS) control architecture is the tuning of static controllers provided for the original formulation, due to the high price volatility. In my research, the objective was to investigate optimization-based adaptive tuning strategies, to exploit both feedback SLS principle and potential available information related to price evolution over the recent past. Two different macro-approaches for the adaptive controller tuning have been derived in my work, whose main difference is represented by how past price information are used for the controller optimization:  ;

;
1. ‘‘Non-parametric optimization’’: find the optimal controllers that maximize final cumulative gain over a past price moving window, up to last available price information.
;
2. ‘‘Parametric optimization’’: identification of a suitable feature set extracted over a past price moving window, provided as input to a Machine Learning Classifier that adaptively updates controller tuning.
;

; Back-testing performed on real-world quotations demonstrates performance improvement introduced by the proposed approaches compared to classical SLS architecture and several benchmark strategies from financial practice. Future work may focus on the investigation of further cost functions for controller optimization considering not only final cumulative gain, but also some indicators of investment risk (e.g., Sortino/Sharpe Ratio).
Reactive trading theory discussed so far relies on the hypothesis of single stock trading. In quantitative finance, optimal allocation of multiple assets in a portfolio represents a well-established problem. Because of the highly non-stationary price dynamics, one of the major issues concerns the identification of a model of returns yielding accurate future predictions. In my research work, I formulated a learning-based Model Predictive Control approach for multi-period portfolio optimization based on a novel “trading-oriented” identification paradigm. Specifically, I showed that allocation performance for multi-asset portfolios can be largely enhanced if the model of returns is computed by maximizing investor’s utility function instead of minimizing prediction error variance. Back-testing conducted on real market data reveals the advantages of the proposed approach against traditional trading strategies. Future work may concern an extensive strategy validation over different portfolios including shares from different markets.