Recipe for Quantitative Trading with Machine Learning

Abstract

On one hand, financial time series are multifractal, thus exhibiting non-Gaussian distribution, the presence of extreme values (outliers), and long-range dependent dynamics. On the other hand, machine learning (ML) models are processes relying heavily on statistical models and methodologies, but treated as black box models due to their inability to explicitly know the relations established between explanatory variables (input) and dependent variables (output). However, when forecasting market returns, or generating autonomous patterns, it is crucial to know the statistical properties of the time series produced by the ML model under consideration. We first present some results on using recurrent neural networks to forecast market returns. We then discuss forecasting stock price directions and propose an algorithm for simultaneously forecasting both future price levels and asset price directions. We also discuss autonomous pattern generation and focus on learning traditional shapes defined by stock analysts. Then, we present complex network analysis to predict stock price fluctuation patterns. At last, we take into consideration the statistical characteristics of financial time series and present a recipe for forecasting both market returns and their directions. We choose to reverse the causality and propose a solution consisting in deciding upon the framework by defining how the model should be specified before beginning to analyse the actual data. We should define the framework by properly formulating model hypotheses which make financial or economic sense, and then carefully determining the number of dependent variables in a regression model, or the number of factors and components in a stochastic model. Based on these observations, we use the multifractal formalism (MF) as a framework for testing the capacity of an ML model to reproduce some non-overlapping statistical properties of the time series. First we define a set of theoretical models with distinct statistical characteristics as well as a set of ML models with the ability to reproduce the properties of the models in the former set. We then train each individual ML model to replicate their respective theoretical model and use ensemble methods to combine these now calibrated ML models to form a meta-model. However, the weights of the latter are statistically computed as a function of the Hurst exponent. As a result, the meta-model is dynamically recombined based on the changing properties of the financial series over time. Once we have identified an ensemble of ML models with specific non-overlapping statistical properties, we can train each constituent model to learn a large number of patterns or technical indicators. We then devise a trading algorithm with strategies accounting for a specific level of Hurst exponent.