In finance, pairs trading is a market strategy that profits from two correlated assets by betting on the mean-reversion of their spread. Thus, the best profits are made by an underpriced and an overpriced asset which revert to their correct price after a short time. Modern technologies allow pairs trading to look for profitable pairs through statistical tests on a custom universe of assets, rather than limiting the research to those that present industrial similarities. However, removing this limitation comes with an exponential increase in the number of potential candidates: in order to overcome the issue, Machine Learning techniques, such as Principal Component Analysis and clustering, are employed to reduce the pairs research to subsets of assets. In this thesis, we start by applying a state-of-the-art pairs selection strategy to the context of the S&P500 daily dataset. We test the procedure with the suggested hyperparameters and evaluate the correlation between a newly proposed measure of cluster consistency and the selected pairs. After that, we assess the sensitivity of the method to the hyperparameters' variation and identify a subset of highly influential ones. Finally, they are optimised through the application of Online Gradient Descent with Momentum: the method achieves good performance with respect to both the suggested static parameters and the average of the tested parametrizations in a transaction fees-free environment, while profits heavily decline when costs are taken into consideration.
In finanza, il trading di coppia è una strategia che guadagna da due cespiti scommettendo sul ritorno alla media della serie numerica data dalla differenza dei loro valori. Le nuove tecnologie permettono al trading di coppia di cercare coppie redditizie attraverso test statistici al posto di somiglianze industriali. La rimozione di tali limitazioni, tuttavia, comporta un aumento esponenziale del numero di potenziali candidati: per superare tale problema, tecniche di apprendimento automatico, come PCA o l'analisi dei gruppi, vengono adottate per ridurre la ricerca delle coppie all'interno di sottoinsiemi di cespiti. In questa tesi, cominciamo con l'applicare, al contesto dei dati giornalieri di S&P 500, una strategia allo stato dell'arte di selezione delle coppie. Testiamo la procedure con gli iperparametri consigliati e valutiamo la correlazione tra una misura di coerenza dei gruppi da noi introdotta e le coppie selezionate. Successivamente, esaminiamo la sensibilità agli iperparametri del metodo ed ne identifichiamo un sottoinsieme di influenti. Infine, li ottimizziamo attraverso l'applicazione di Online Gradient Descent with Momentum: in un ambiente senza costi di transazione, il metodo ottiene buone prestazioni sia rispetto ai parametri statici suggeriti che alla media delle parametrizzazioni testate, mentre i guadagni diminuiscono quando i costi vengono considerati.
Hyperparameter tuning for pairs trading : an online learning approach
COLIVA, NAHUEL
2021/2022
Abstract
In finance, pairs trading is a market strategy that profits from two correlated assets by betting on the mean-reversion of their spread. Thus, the best profits are made by an underpriced and an overpriced asset which revert to their correct price after a short time. Modern technologies allow pairs trading to look for profitable pairs through statistical tests on a custom universe of assets, rather than limiting the research to those that present industrial similarities. However, removing this limitation comes with an exponential increase in the number of potential candidates: in order to overcome the issue, Machine Learning techniques, such as Principal Component Analysis and clustering, are employed to reduce the pairs research to subsets of assets. In this thesis, we start by applying a state-of-the-art pairs selection strategy to the context of the S&P500 daily dataset. We test the procedure with the suggested hyperparameters and evaluate the correlation between a newly proposed measure of cluster consistency and the selected pairs. After that, we assess the sensitivity of the method to the hyperparameters' variation and identify a subset of highly influential ones. Finally, they are optimised through the application of Online Gradient Descent with Momentum: the method achieves good performance with respect to both the suggested static parameters and the average of the tested parametrizations in a transaction fees-free environment, while profits heavily decline when costs are taken into consideration.File | Dimensione | Formato | |
---|---|---|---|
2022_07_Coliva_01.pdf
accessibile in internet per tutti
Descrizione: Testo della tesi
Dimensione
5.37 MB
Formato
Adobe PDF
|
5.37 MB | Adobe PDF | Visualizza/Apri |
2022_07_Coliva_02.pdf
accessibile in internet per tutti
Descrizione: Executive summary
Dimensione
4.98 MB
Formato
Adobe PDF
|
4.98 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/10589/189963