Nikolay Nikolaev, Goldsmiths College, University of London

Automated Algorithmic Trading

NDS

Neural Dynamical Systems

The development of real-time trading algorithms operating on high-frequency data depends crucially on the ability to capture the temporally changing character in financial time series, known as nonstationarity. Special attention is devoted to learning time-dependent statistical characteristics of the series (mean and variance) through dynamical state-space models represented using nonlinear neural networks. Recent research implemented a library of algorithms for sequential Bayesian filtering for nonlinear Neural Dynamical Systems (NDS). Following this methodology helps us to handle properly heavy tailed and skewed distributions so as to make probabilistic predictions, that is to produce esimates of the future data along with confidence intervals for the belief in them. Another advantage of NDS is that they can be trained directly to maximise the ratio beween the first and second moments of the returns so as to achive risk-adjusted performance.

The neural network algorithms are suitable for algorithmic trading with intra-day "tick-by-tick" data. They generate online buy/sell signals (in milliseconds) and facilitate placing instantaneous orders in the market. Our current work tailors trading algorithms for such applications like:

- stock price forecasting (and stock market index modelling);

- bond price forecasting (and interest-rate modelling);

- currency price forecasting (and currency modelling).

The NNDS algorithms show high success rates, positive profit and loss, as well as maximised Sharpe ratios. These qualities makes them useful for various participants in the electronic financial markets such as: market makers, brokers, hedge funds, and institutional traders.

PKM

Probabilistic Kernel Machines

A contemporary alternative to tackling nonstationary behaviour in time series is provided by the Probabilistic Kernel Machines (PKM). These are predictive tools whose parameters are analytically tractable and can be obtained with sequential training algorithms when financial data are processed. Applying hierarchical Bayesian inference to the kernel machines allows us to include heavy tail (Student-t) noise hyperparameters, which are updated along with the arrival of the data in time so as to achieve efficient modelling of nonstationary time series. From a practical point of view, this enables us also to deal with extreme shocks in the observations, like outliers and breaks, as well as to detect change points in the series. From a theoretical point of view, this also helps to examine the robustness of the learned models by computing recursive residuals and estimating such characteristics like influences and leverages.

The probabilistic kernel machines, like SVM, RVM, KRLS, are appropriate for electronic trading with daily data. Our experience includes design of customized systems for statistical trading in the following application areas:

- trading of stock market indices (and stock market index modelling);

- forecasting electricity prices (and electricity demand);

- cross-currency exchange rates prediction (and FX modelling).

PKM are found successfull in automatic adjustment of hedging positions, in response to the flow of information, and generation of orders to position risks. PKM are trend following algorithms that can be used at medium to high frequency, that is they can be used for taking hundreds of decisions per day.