Publication: Operando en el mercado de valores: análisis financieros híbridos y computación evolutiva
Loading...
Official URL
Full text at PDC
Publication Date
2015-02-06
Authors
Advisors (or tutors)
Editors
Journal Title
Journal ISSN
Volume Title
Publisher
Universidad Complutense de Madrid
Citations
Exportar
Abstract
Esta tesis presenta la implementación de un innovador sistema de comercio automatizado que utiliza tres importantes análisis para determinar lugares y momentos de inversión. Para ello, este trabajo profundiza en sistemas automáticos de comercio y estudia series temporales de precios históricos pertenecientes a empresas que cotizan en el mercado bursátil. Estudiamos y clasifcamos las series temporales mediante el uso de una novedosa metodología basada en compresores de software. Este nuevo enfoque permite un estudio teórico de la formación de precios que demuestra resultados de divergencia entre precios reales de mercado y precios modelados mediante paseos aleatorios, apoyando así el desarrollo de modelos predictivos basados en el análisis de patrones históricos como los descritos en este documento. Además, esta metodología nos permite estudiar el comportamiento de series temporales de precios históricos en distintos sectores industriales mediante la búsqueda de patrones en empresas pertenecientes al mismo sector. Los resultados muestran agrupaciones que indican tendencias de mercado compartidas y ,por tanto, señalan que la inclusión de un análisis industrial puede reportar ventajas en la toma de decisiones de inversión. Comprobada la factibilidad de un sistema de predicción basado en series temporales y demostrada la existencia de tendencias macroeconómicas en las diferentes industrias, proponemos el desarrollo del sistema completo a través de diferentes etapas. Iterativamente y mediante varias aproximaciones, testeamos y analizamos las piezas que componen el sistema nal. Las primeras fases describen un sistema de comercio automatizado, basado en análisis técnico y fundamental de empresas, que presenta altos rendimientos y reduce el riesgo de pérdidas. El sistema utiliza un motor de optimización guiado por una versión modi cada de un algoritmo genético el la que presentamos operadores innovadores que proporcionan mecanismos para evitar una convergencia prematura del algoritmo y mejorar los resultados de rendimiento nales. Utilizando este mismo sistema de comercio automático proponemos técnicas de optimización novedosas en relación a uno de los problemas más característicos de estos sistemas, el tiempo de ejecución. Presentamos la paralelización del sistema de comercio automatizado mediante dos técnicas de computación paralela, computación distribuida y procesamiento grá co. Ambas arquitecturas presentan aceleraciones elevadas alcanzando los x50 y x256 respectivamente. Estápas posteriores presentan un cambio de metodologia de optimización, algoritmos genéticos por evolución gramatical, que nos permite comparar ambas estrategias e implementar características más avanzadas como reglas más complejas o la auto-generación de nuevos indicadores técnicos. Testearemos, con datos nancieros recientes, varios sistemas de comercio basados en diferentes funciones de aptitud, incluyendo una innovadora versión multi-objetivo, que nos permitirán analizar las ventajas de cada función de aptitud. Finalmente, describimos y testeamos la metodología del sistema de comercio automatizado basado en una doble capa de gramáticas evolutivas y que combina un análisis técnico, fundamental y macroeconómico en un análisis top-down híbrido. Los resultados obtenidos muestran rendimientos medios del 30% con muy pocas operaciones de perdidas.
This thesis concerns to the implementation of a complex and pioneering automated trading system which uses three critical analysis to determine time-decisions and portfolios for investments. To this end, this work delves into automated trading systems and studies time series of historical prices related to companies listed in stock markets. Time series are studied using a novel methodology based on clusterings by software compressors. This new approach allows a theoretical study of price formation which shows results of divergence between market prices and prices modelled by random walks, thus supporting the implementation of predictive models based on the analysis of historical patterns. Furthermore, this methodology also provides us the tool to study behaviours of time series of historical prices from di erent industrial sectors seeking patterns among companies in the same industry. Results show clusters of companies pointing out market trends among companies developing similar activities, and suggesting a macroeconomic analysis to take advantage of investment decisions. Tested the feasibility of prediction systems based on analyses related to time series of historical prices and tested the existence of macroeconomic trends in the industries, we propose the implementation of a hybrid automated trading system through several stages which iteratively describe and test the components of the nal system. In the early stages, we implement an automated trading system based on technical and fundamental analysis of companies, it presents high returns and reducing losses. The implementation uses a methodology guided by a modi ed version of a genetic algorithm which presents novel genetic operators avoiding the premature convergence and improving nal results. Using the same automated trading system we propose novel optimization techniques related to one of the characteristic problems of these systems: the execution time. We present the parallelisation of the system using two parallel computing techniques, rst using distributed computation and, second, implementing a version for graphics processors. Both architectures achieve high speed-ups, reaching 50x and 256x respectively, thus, they present the necessary speed-ups required by systems analysing huge amount of nancial data. Subsequent stages present a transformation in the methodology, genetic algorithms for grammatical evolution, which allows us to compare the two evolutionary strategies and to implement more advanced features such as more complex rules or the self-generation of new technical indicators. In this context, we describe several automated trading system versions guided by di erent tness functions, including an innovative multi-objective version that we test with recent nancial data analysing the advantages of each tness function. Finally, we describe and test the methodology of an automated trading system based on a double layer of grammatical evolution combining technical, fundamental and macroeconomic analysis on a hybrid topdown analysis. The results show average returns of 30% with low number of negative operations.
This thesis concerns to the implementation of a complex and pioneering automated trading system which uses three critical analysis to determine time-decisions and portfolios for investments. To this end, this work delves into automated trading systems and studies time series of historical prices related to companies listed in stock markets. Time series are studied using a novel methodology based on clusterings by software compressors. This new approach allows a theoretical study of price formation which shows results of divergence between market prices and prices modelled by random walks, thus supporting the implementation of predictive models based on the analysis of historical patterns. Furthermore, this methodology also provides us the tool to study behaviours of time series of historical prices from di erent industrial sectors seeking patterns among companies in the same industry. Results show clusters of companies pointing out market trends among companies developing similar activities, and suggesting a macroeconomic analysis to take advantage of investment decisions. Tested the feasibility of prediction systems based on analyses related to time series of historical prices and tested the existence of macroeconomic trends in the industries, we propose the implementation of a hybrid automated trading system through several stages which iteratively describe and test the components of the nal system. In the early stages, we implement an automated trading system based on technical and fundamental analysis of companies, it presents high returns and reducing losses. The implementation uses a methodology guided by a modi ed version of a genetic algorithm which presents novel genetic operators avoiding the premature convergence and improving nal results. Using the same automated trading system we propose novel optimization techniques related to one of the characteristic problems of these systems: the execution time. We present the parallelisation of the system using two parallel computing techniques, rst using distributed computation and, second, implementing a version for graphics processors. Both architectures achieve high speed-ups, reaching 50x and 256x respectively, thus, they present the necessary speed-ups required by systems analysing huge amount of nancial data. Subsequent stages present a transformation in the methodology, genetic algorithms for grammatical evolution, which allows us to compare the two evolutionary strategies and to implement more advanced features such as more complex rules or the self-generation of new technical indicators. In this context, we describe several automated trading system versions guided by di erent tness functions, including an innovative multi-objective version that we test with recent nancial data analysing the advantages of each tness function. Finally, we describe and test the methodology of an automated trading system based on a double layer of grammatical evolution combining technical, fundamental and macroeconomic analysis on a hybrid topdown analysis. The results show average returns of 30% with low number of negative operations.
Description
Tesis inédita de la Universidad Complutense de Madrid, Facultad de Informática, Departamento de Arquitectura de Computadores y Automática, leída el 02-07-2014