The conference is going to take place at Eötvös Loránd University, Budapest, Hungary Pázmány Péter Sétány 1/A, Harmony Hall.
Using a large-scale Deep Learning approach applied to a high-frequency database containing billions of electronic market quotes and transactions for US equities, we uncover nonparametric evidence for the existence of a universal and stationary price formation mechanism relating the dynamics of supply and demand for a stock, as revealed through the order book, to subsequent variations in its market price. We assess the model by testing its out-of-sample predictions for the direction of price moves given the history of price and order flow, across a wide range of stocks and time periods. The universal price formation model is shown to exhibit a remarkably stable out-of-sample prediction accuracy across time, for a wide range of stocks from different sectors. Interestingly, these results also hold for stocks which are not part of the training sample, showing that the relations captured by the model are universal and not asset-specific.
The universal model – trained on data from all stocks – outperforms, in terms of out-of-sample prediction accuracy, asset-specific linear and nonlinear models trained on time series of any given stock, showing that the universal nature of price formation weighs in favour of pooling together financial data from various stocks, rather than designing asset or sector-specific models as commonly done. Standard data normalizations based on volatility, price level or average spread, or partitioning the training data into sectors or categories such as large/small tick stocks, do not improve training results. On the other hand, inclusion of price and order flow history over many past observations is shown to improve forecasting performance, showing evidence of path-dependence in price dynamics.
Automated digital consultancy platforms (“Robot advisors”) reduce costs and improve the quality of the perceived service, speeding it up and making user involvement more transparent. These improvements are often offset by risk classification models that are simpler than those employed in traditional consultancy. We aim to exploit the large amount of data available to robot advisors to build portfolios that are better fit to the risk profiles of the investors. This is made possible, on one hand, constructing groups of homogeneous risk profiles, based on user responses to MIFID questionnaires and, on the other hand, constructing homogeneous clusters of financial assets, based on their risk and return performance. We demonstrate that machine learning methods and, specifically, network models, can be used to “automatise” the previous classifications and, eventually, to assess whether an investor’s portfolio matches its risk profile. We will apply the proposed methodology to “classical” portfolios, based on exchange traded funds, and to “hybrid” portfolio that contain bitcoin assets, traded in different exchange markets, for which there is the additional issue of finding which market drives the price.
Government bond trading is going through unprecedented change today, the role of traditional (mostly voice broking) bond trading model is decreasing due to the electrification of the markets. The electrification means more reliable and good quality data that give a solid foundation to the application of advanced statistical approaches.
Our aim is to highlight applications in bond trading where machine learning techniques offer improvements.
Recently more and more investors have been extracting information from news articles automatically in order to modify their investment decisions. The usage of similar techniques for risk management is less usual. Our goal is to show how automated text analysis can be tailored to extract credit-related information and how such a system can make the work of credit coverage analysts easier and more efficient.
We have studied the predictability of different range-based stock volatility estimators. Our aim was to compare the accuracies of a simple direction-of-change forecasting framework applied to different volatility estimates. Recurrent neural networks were applied to make the predictions. We present empirical results for all 30 constituents of the Dow Jones Industrial Average index. According to our results, range-based volatility estimates can easily be predicted to some degree of accuracy, while the so popular close-to-close estimator (the standard deviation of daily log returns) seemed essentially unpredictable.
We explain two successful recent applications of machine learning techniques in mathematical Finance, namely learning algorithms for calibration functionals and solving a real world risk management problem (based on joint works with Hans Buehler, Christa Cuchiero, Lukas Gonon, Wahid Khosrawi-Sardroudi, and Ben Wood). Several outlooks towards new directions are also presented.
This talk provides a survey of discrete time, multi period, sequential investment strategies for financial markets. Under memoryless assumption on the underlying process generating the asset prices, the log-optimal portfolio achieves the maximal asymptotic average growth rate. For general dynamic portfolio selection, when asset prices are generated by a stationary and ergodic process, growth optimal empirical strategies are shown, where some principles of nonparametric regression estimation and of machine learning aggregation are applied. The empirical performance of the methods is illustrated for NYSE data.
Calibration time being the bottleneck for models with rough volatility, we present ways for substantial speed-ups, along every step of the calibration process: In a first step we describe a powerful numerical scheme (based on functional central limit theorems) for pricing a large family of rough volatility models. In a second step we discuss various machine learning methods that significantly reduce calibration time for these models. By simultaneously calibrating several (classical and rough) models to market data, we re-confirm as a byproduct of our calibration results, that volatility is rough, calibration performance being best for very small Hurst parameters in a multitude of market scenarios.
An algebraic approach for derivatives pricing using feedforward neural networks is presented. The approach is characterized by fast execution speed and better generalization properties than contemporary optimization techniques. Potential application to dynamic initial margin modelling is discussed.
Tensorflow, Google’s machine learning package has been developed for deep neural network modeling. Less known is, however, that the platform brings great features for derivatives pricing too given its embedded AD (automatic differentiation) capabilities as will be illustrated by a couple of examples in the talk.
email: email@example.comThe main sponsor is