Collaborative filtering ap- proaches proved to be effective for recommender systems in predicting user preferences using past known user ratings of items. The first approach is theory-based and attempts to find a numerical solution to the problem by approximating the true solution of the boundary value problem, containing the constitutive equations, by using a simple feedforward neural network. This possibility opened a great number of new research areas for statisticians and computer scientists. We study learned image compression at variable bitrates in the context of subsequent classification, where the classifiers are parameterized as convolutional neural networks. An important conclusion is that the methods based on marginal Cox estimates work well if the marginal hazard ratio is estimated on a time interval in the beginning of the study and for a suitable interval length. Moreover, its performance does not depend on additional tuning parameters.

This thesis presents the theory and main ideas behind some of the nowadays most popular methods used for causal structure learning as well as the ICP algorithm, a new algorithm based on a method recently developed at ETH Zurich. While this version yields relatively good results, in some cases the PC algo- rithm eliminates a large part of true positive edges from the search space. In the end, the Additivity and Variance Stabilizing transformation is presented and applied in datasets. Different estimators based on the ideas of neighborhood selection and algorithms to compute them are presented. However, a proportional hazards model can only be reformulated as an accelerated failure time model, if the underlying distribution of the survival times is Weibull Cox and Oakes, , Section 5. We found that on data sets with only few variables, the procedure performs satisfactorily, which means that the underlying data generation process is well reected. In contrast to conventional methods like lasso or ridge regression, this method is able to discover signals, which are neither sparse nor dense.

Jobs: SwissQuant Group – 15 aktuelle Stellen bei SwissQuant Group | Jooble

For variable selection, we study a newly proposed and very interesting idea called knockoffs. We investigate a form of time series model training that coincides with the forecasting procedure, and compare it to the standard modes of training. Instead we propose two contrasting approaches: We propose two estimation approaches that directly build up on their regression estimator and a third procedure which is analogous to their regression estimator, but modified to match the likelihood arising in the case of covariance matrices.


In ordinal classification, we tried three types of methods: Experimental evaluations of the proposed algorithms for various real life applications show a potential to improve the opti- mization performance of the network. In classical literature, for example in G.

A rather general linear model with hidden variables is introduced and an invariant causal prediction framework for such models is established.

master thesis at swissquant

In this thesis we review three fundamental asymptotic results for empirical processes, as presented in van der Waart and Wellner As a consequence the application of some statistical techniques to such problems is slowed down considerably from a computational point of view by the high dimensionality of the data: Marloes Henriette Maathuis Jul Abstract: The Log-Periodic Power Law Singularity LPPLS model is an attempt to model unsustainable growth in financial markets, namely super-exponential growth, and predict tgesis in- evitable burst of such bubbles.

In the case of RNNs, our training method is empirically shown to be more a to prediction mistakes when performing forecasts in comparison to RNNs trained via the conditional maximum likelihood.

Navigation Area

It is also important to point out that a deep comparison between the models is not completely fair, due to the their inner nature: The second estimates the ocurrence of a deforestation event by comparing the last few predicted and real reflectance values. Maathuis Leonard Henckel Oct Abstract: The results we obtained here support and strengthen our previous findings. Look no further and apply now! This thesis presents the theory and main ideas behind some of the nowadays most popular methods used for causal structure learning as well as the ICP algorithm, a new algorithm based on a method recently developed at ETH Zurich.

As this work is the first analysis of such kind of this dataset, there are manifold extension possibilities such as applying dynamic topic modeling to retrieve the evolution of topics. Friedman’s gradient boosting machine is a stage-wise additive algorithm, where the loss of the model is minimized by adding weak learners in a gradient-descent manner.

master thesis at swissquant

Moreover sometimes it ghesis even be costly to store the data itself. Our goal is to train a compression al- gorithm, based on recurrent neural networks, such that the accuracy of pretrained classifiers which are unknown to the compression system, is maintained, when evaluated on the compressed dataset, also at low bitrates.


The proposed approach consists of a hie-rarchical convolutional neural network CNN that models the inherent structure of documents. These recommender systems aim to maximise the acceptance rate of clients agreeing to buy the recommended instrument.

We compare and test the sample properties of the algorithms on simulated data. The performance of each recommender system is evaluated using four different ranking metrics: Given a spatial dataset, consider a nonparametric regression model where the aim is to estimate the regression surface.

swissQuant Master Thesis Job in Z├╝rich | Glassdoor

However, identifying causal genes in GWAS studies remains challenging because many statistically significant variants are in linkage disequilibrium with causal variants and some loci harbor more than one causal variant. From a Natural Language Processing perspective, this work investigates the role that information over the human way of processing can have in augmenting classical sentiment classifiers.

And hierarchical testing exploits the correlation structure of the variables to adapt its resolution level to the strength of the signal in a data-driven manner. Moreover, this thesis analyzes how well the issue of predicting parking lot occupancy rates can be ar to other cities.

Moreover, we analyse the causal dependencies of the variables used for the pre- diction models.

swissQuant Group AG

Graphical models can answer the thrsis whether a causal relation exists in an obser-vational design. Moreover, we will try to identify any hidden evolving pattern over the years, which can be seen only using some dynamic linear model, thanks to the famous state space representation, which will be finally compared to the outcomes of the other more classical approaches introduced before.

The aim of this thesis is to define and implement a framework for event swisquant and characterization on Twitter streams using unsupervised statistical techniques. In this thesis, an approach already successfully applied to determine the dimension of mixtures of mutagenetic trees MMT by Yin et al.