Deep learning neural networks based algorithmic trading strategy for colombian financial market using tick by tick and order book data

This work presents an innovative and highly competitive Algorithmic Trading (AT) Strategy, based on a Convolutional Neural Network price direction predictor that uses High Frequency (HF) transactions and Limit Order Book (LOB) data. Information used includes data from US and Colombian market. Data processing include more than 5 million raw data files of 21 stocks from different industries (Energy, Finance, Technology, Construction, among others). Since data include two different sources (Transaction and LOB), applying feature engineering is necessary to homogenize inputs. For transaction data, an image-like representation (Grammian Angular Field GAF) is used. It converts Financial Time Series (FTS) to polar coordinates and creates a kernel based on cosine differences. Additionally, this work proposes a transformation for LOB data. This representation includes all available information deviated from LOB raw data and it will create an image-like representation of LOB. These two sources will feed up into a proposed 3D-Convolutional Neural Network (3D-CNN) architecture that generates price direction predictions. These predictions will serve as a trading signal generator for two Algorithmic Trading Strategies. Both of them take real market constrains into consideration, such as liquidity provision, transaction costs, among others. The two proposed strategies works under different risk aversion constrains. Results from the proposed 3D-CNN predictor present a strong performance, ranging between 70% and 74% in Directional Accuracy (DA), while reducing model parameters as well as making inputs time invariant. Moreover, trading strategies results illustrate that the proposed CNN predictor can lead to profitable trades and liquidity improvement in the Colombian Market. Testing results for both AT strategies on Colombian Market Data lead to interesting findings. Under different constrains of take profit, stop loss and transaction cost, both strategies aggressive and conservative lead to positive returns over the same period of time. Moreover, results of number of trades performed by the aggressive AT helps to understand how AT may impact positively liquidity provision in developing financial markets.

• Autores: Niño Peña, Jaime Humberto
• Tipo de recurso: Doctoral thesis
• Fecha de publicación: 2019
• Institución: Universidad Nacional de Colombia
• Repositorio: Universidad Nacional de Colombia
• Idioma: spa
• OAI Identifier: oai:repositorio.unal.edu.co:unal/69861
• Acceso en línea: https://repositorio.unal.edu.co/handle/unal/69861; http://bdigital.unal.edu.co/72209/
• Palabra clave: 0 Generalidades / Computer science, information and general works; 02 Bibliotecología y ciencias de la información / Library and information sciences; 33 Economía / Economics; 5 Ciencias naturales y matemáticas / Science; 6 Tecnología (ciencias aplicadas) / Technology; 62 Ingeniería y operaciones afines / Engineering; Algorithmic Trading; Deep Learning; Convolutional Neural Networks; Computational Finance; High Frequency Trading; Financial Time Series; Finanzas computacionales; Aprendizaje profundo; Redes convolucionales; Aprendizaje de representación; Libro de órdenes; Transacciones; Estrategias de Negociación algorítmica; Negociación de alta frecuencia
• Rights: openAccess
• License: Atribución-NoComercial 4.0 Internacional