Dienstag, 15.10.2019, 12-13 Uhr - Raum: W9-109
Lennart Oelschläger, M.Sc.
Universität Bielefeld
Detection of Bearish and Bullish Markets in the DAX Using Hierarchical Hidden Markov Models
Almost any financial market exhibits alternating periods of rising and falling prices. Stock traders have to mind those trends to make profitable investment decisions. So far and among other models, the hidden Markov model has been used to quantify the trend changes in order to predict upcoming market behaviour. However, this model in
its basic form is not capable of capturing short-term and long-term trends jointly. Extending the model by a hierarchical structure, we aim to fix this deficit and to draw a more comprehensive picture of the financial market. The improvement is exemplary investigated for the Deutscher Aktienindex.
The talk will be held in English.
Dienstag, 29.10.2019,12-13 Uhr - Raum: W9-109
Anna-Kaisa Ylitalo, Ph.D.
Luke (Natural Resources Institute Finland)
Analysis of wolf movements during pup-rearing season
Movement behaviour of wolves (Canis lupus) have a special form in the summer, when they have pups in a den to be fed. We analysed preying trips of nine radio-collared wolves tracked in Finland in 2006-2013. According to the spatial position observations and a trip time covariate, the movement track of a wolf was split into segments using hidden Markov models. Those segments are considered to represent different kind of movement modes, namely resting, moving and return. The resulting segmentation of the movement is studied further using an additional data set consisting of detailed field information on den sites and prey.
The talk will be held in English.
Dienstag, 12.11.2019, 12-13 Uhr - Raum: W9-109
Daniel Engler, M.Sc.
Universität Kassel
Identity strength and priming effects
Using a stated choice experiment, we find that a prime that makes environmental identity salient makes people behave greener. Furthermore, we discover non-linear priming effects for environmental identity, which means that raising the salience of highly environmentally oriented respondents or respondents without environmental identity does not change behav-ior while it does for respondents with a medium level strength of identity. Methodologically, our study combines for the first time a priming experiment with a stated choice (SC) exper-iment and uses a respondent specific status quo alternative in the empirical analysis with mixed logit models.
Dienstag, 26.11.2019, 12-13 Uhr - Raum: W9-109
Kaja Balzereit, M. Sc.
Fraunhofer IOSB-INA
Analyzing Industrial Production Plants using Machine Learning Methods
Modern industrial production plants contain numerous sensors which generate large datasets. Analysis of these datasets allows detection of anomalies in the plant behavior or varying product quality. However, analyzing these datasets poses great challenges, especially since such data contains continuous and discrete values, and thus needs to be analyzed by hybrid methods. This talk presents a hybrid machine learning method based on two different conceptual layers (sub-symbolic and symbolic), allowing detection of deviations in the plant behavior by intelligent data analysis.
The talk will be held in English.
Dienstag, 10.12.2019, 12-13 Uhr - Raum: W9-109
Dr. Francisco J. Bahamonde-Birke
Utrecht University
Establishing the variability of multinomial discrete models. Which proportion of the variability is actually being explained by our models?
As discrete choice models do not allow measuring accurately the model’s error, it is not possible to construct a direct equivalent to the coefficient of determination. Several pseudo-equivalents for the coefficient of determination have been proposed in the past, but they do not aim at establishing the model’s explained variability akin to the coefficient of determination. This paper offers an extensive discussion on methods to compute the adjustment of multinomial choice models. Itfurther proposes a method to estimate the total variability of the model at the level of the utility functions accurately and proposes an index that preserves the properties of in multiple dimensions. Furthermore, the method allows assessing the relative importance of a given explanatory variable on the model at aggregated level or to measure the impact of random error terms on the model’s variability.
The talk will be held in English.
Dienstag, 07.01.2020, 12-13 Uhr - Raum: W9-109
Dr. Emmeke Aarts
Utrecht University
Mixed hidden Markov models using Bayesian estimation: the R package mHMMbayes
The mixed hidden Markov model (HMM) is a generalization of the well-known hidden Markov model, tailored to accommodate (intense) sequential data of multiple individuals or animals simultaneously. Using a mixed effects framework, we allow for heterogeneity in the model parameters (transition probability matrix and conditional distribution), while estimating one overall HMM. The package mHMMbayes a useful tool to estimate random effects HMMs in the programming language R, and the only CRAN package that allows fitting such and fixed Bayesian HMM models. The model can be fitted on multivariate data with a categorical distribution, and include individual level covariates (allowing for e.g., group comparisons on model parameters). Parameters are estimated using Bayesian estimation utilizing the forward-backward recursion within a hybrid Metropolis within Gibbs sampler. The package also includes various automated visualizations of the fitted model, a function to simulate data, and a function to obtain the most likely hidden state sequence for each individual using the Viterbi algorithm.
The talk will be held in English.
Dienstag, 21.01.2020, 12-13 Uhr - Raum: W9-109
Aktuelle Forschungsthemen im ZeSt