Abstract: One of the main components of optimally scheduling portfolio trading is the estimation of intraday and multi-day propagation of risk and market impact. This estimation is dependent on accurately modelling the dynamics involved across multiple scales: ranging from the daily to the tick scale.
This talk is concerned with how machine learning can be used to inform those dynamical models, exploring a range of techniques, including LSTMs, CNNs, Bayesian inference and Clustering methods.
One of our examples is concerned with the real-time prediction of liquidity, as it is, in general, one of the most important inputs to trading models. Observing that liquidity shocks, which can be both positive and negative, can be significant, we create a framework that can update the models’ predictions for those shocks for different horizons. Utilising that framework, we test Classical, Bayesian as well as Neural Network techniques, where we find that Neural Network architectures can indeed provide an edge in real-time liquidity prediction.
Beyond predicting market variables such as liquidity, we also demonstrate that, by using unsupervised learning techniques, such as clustering, on market microstructure features, we can better identify homogeneous groups among stocks. There are many applications of these stock clusters, including being a quick proxy of trading difficulty as well as an anomaly detection tool. Most importantly, however, they allow us to divide the global trading universe before calibrating cluster-specific trading models, while moving away from typical country and sector groupings. This not only helps in improving model calibration, but it also provides new insights regarding intraday trading phenomena.
Bio: Andreas is working as a Quantitative Researcher at Goldman Sachs Quantitative Execution Services, with an emphasis in machine learning techniques for execution algorithms. Andreas has received a PhD in Information Engineering at the University of Cambridge, focusing on the interface of stochastic control theory and Bayesian machine learning, where he developed graph theoretic tools for predicting the shapes of the probability distributions to arise in the observable time-series due to the underlying non-linear stochastic interconnections. Andreas’ teaching experience included several engineering undergraduate courses, including Inference and Machine Learning, Linear Algebra, Probability, Control and Signal Processing. During his PhD, he has also worked at Informetis Europe as a Machine Learning Algorithm Engineer, developing efficient Bayesian inference techniques for smart electricity meter applications. Andreas also holds a BA and an MEng degree in Electrical and Information Sciences from Trinity College, University of Cambridge, during which he has received the G-Research and The Technology Partnership (TTP) awards, while his Master’s thesis was done in collaboration with British Cycling, developing a racing cyclist fitness predictor.