|
|
Modeling Dynamics of Correlations between Stock Markets with High-frequency Data
Lypko, Vyacheslav ; Baruník, Jozef (advisor) ; Krištoufek, Ladislav (referee)
In this thesis we focus on modelling correlation between selected stock markets using high-frequency data. We use time-series of returns of following indices: FTSE, DAX PX and S&P, and Gold and Oil commodity futures. In the first part of our empirical work we compute daily realized correlations between returns of subject instruments and discuss the dynamics of it. We then compute unconditional correlations based on average daily realized correlations and using feedforward neural network (FFNN) to assess how well the FFNN approximates realized correlations. We also forecast daily realized correlations of FTSE:DAX and S&P:Oil pairs using heterogeneous autoregressive model (HAR), autoregressive model of order p (AR(p)) and nonlinear autoregressive neural network (NARNET) and compare performance of these models.
|
|
|
Alternative field curve modelling approach : regional models
Šopov, Boril ; Seidler, Jakub (advisor) ; Baruník, Jozef (referee)
In this thesis, we focus on thorough yield curve modelling. We build on extended classical Nelson-Siegel model, which we further develop to accommodate unobserved regional common factors and principal components. We centre our discussion on central European currencies' yield curves: CZK, HUF, PLN and SKK. We propose two novel models to capture regional dynamics; one based purely on state space formulation and the other relying also on principal components of the regional yield curves. Moreover, we supplement the models with two application examples in risk management and structural break detection. The main contribution of this thesis is a creation of a complete framework that enables us to analyse yield curves, to design risk scenarios and to detect structural breaks of various types.
|
|
|
Nowcasting the Real GDP Growth of the European Economies based on Machine Learning
Baylan, Su Hazal ; Kočenda, Evžen (advisor) ; Baruník, Jozef (referee)
This thesis analyzes the nowcasting of quarterly GDP growth for nine European economies using a dynamic factor model and four different machine learning models. These machine learning models are as follows: Ridge, Lasso, Elastic Net, and Random Forest. The data includes ten hard and fifteen soft indicators for each country in order to calculate GDP for each nowcasting iteration for pre-covid and covid periods. For machine learning, models are fed with the extracted factors that are obtained from the dynamic factor model, and for all nowcasting models expanding window approach is selected to estimate nowcasting iterations. The empirical finding indicates that overall machine learning models provide better forecasting accuracy compared to dynamic factor models and benchmark models for more stable periods, such as the period before Covid-19. On the other hand, for more volatile periods where the uncertainties are higher in economies, the dynamic factor model outperforms machine learning models in order to nowcast GDP growth. In addition to this, Random Forest is able to outperform all the alternative models for small economies such as Slovenia and Portugal for stable periods. JEL Classification C01, C33, C53, C83, E37 Keywords Nowcasting, DFM, Ridge, Lasso, Elastic Net, Random Forest Title Nowcasting...
|
|
|
Price Impact of Order Book Events in Bitcoin Market
Erben, Marek ; Šíla, Jan (advisor) ; Baruník, Jozef (referee)
1 Abstract This thesis examines the price impact of order book events in the Bitcoin mar- ket. Using the data obtained from Binance exchange, the thesis shows that short-term price changes can be explained by high-frequency demand-supply interaction depicted in the Limit Order Book (LOB). The thesis demonstrates that the instantaneous price impact function has a non-linear shape, indicating that small and large orders have di↵erent e↵ects on price, potentially leading to opportunities for price manipulation and quasi-arbitrage. Additionally, the analysis confirms the inverse relation between the price impact coe cient and market depth. Furthermore, the thesis observes that there are no clear intraday patterns for the price impact coe cient. These findings provide valuable insights into the understanding of Bitcoin's price dynamics, benefiting traders, investors, and policymakers seeking to understand the complexities of the cryptocurrency market. 1
|
|
|
Stock Ownership Structure and Related Risk Premium
Rosický, Ondřej ; Baruník, Jozef (advisor) ; Kočenda, Evžen (referee)
Goal of this thesis is to discover the possible risk premium for stocks with respect to their ownership structure. We work with two types of investors, retail and institutional. Those types of investors have different expectations, preferences and behave differently in certain market events. We built the long-short IMR (institutional minus retail) factor as difference in returns of top and bottom portfolios based on proportion of institutional ownership and added this factor to Fama and French Three Factor Model. There is approximately 0.23 % risk premium for stocks with high share of institutional owners. Further we also try to find the possible impact of nominal stock price on ownership structure. With higher nominal price there is higher institutional ownership. On the other hand, this impact is negligible for low and high percentage share of institutional ownership, therefore IMR factor could not be substituted by the nominal stock price. Lastly, we tried to discover what causes the abnormal returns after the execution date. We found out that with increase in retail ownership by 1 p.p., the abnormal returns are higher in one week after stock split execution date by 0.8 p.p. That is in line with earlier discovered risk premium because with the decrease in the portion of institutional ownership...
|
|
|
Binning numerical variables in credit risk models
Mattanelli, Matyáš ; Baruník, Jozef (advisor) ; Teplý, Petr (referee)
This thesis investigates the effect of binning numerical variables on the per- formance of credit risk models. The differences are evaluated utilizing five publicly available data sets, six evaluation metrics, and a rigorous statistical test. The results suggest that the binning transformation has a positive and significant effect on the performance of logistic regression, feedforward artifi- cial neural network, and the Naïve Bayes classifier. The most affected aspect of model performance appears to be its ability to differentiate between eligible and ineligible customers. The obtained evidence is particularly pronounced for moderately-sized data sets. In addition, the findings are robust to the inclusion of missing values, the elimination of outliers, and the exclusion of categorical features. No significant positive effect of the binning transformation was found for the decision tree algorithm and the Random Forest model.
|
|
|
Predicting stock price movements from financial news using deep neural networks
Kramoliš, Richard ; Baruník, Jozef (advisor) ; Vácha, Lukáš (referee)
Financial media are an important source of information and many articles about companies and stocks are released every day. This thesis assesses the informa- tion value of the articles and utilizes these articles for the stock price move- ment prediction task. For this purpose, models with transformer architecture are used, specifically Bidirectional Encoder Representations from Transform- ers. These models are able to process the text data and create the contextual representation of the text sequence. After adding the classification layer, the models are applied for the stock price movement predictions. The thesis evalu- ates multiple models including different techniques and parameters to find the best performing model. It focuses on two data filters that are expected to de- crease the noise in the data. Moreover, it introduces a new method to recognize the company of interest. As a result of the hyperparameter optimization, the final model is constructed. JEL Classification C45, C51, C52, C53, G11, G14, G17 Keywords BERT, Transformer, Financial Articles, Stock Trading Title Predicting stock price movements from financial news using deep neural networks
|
|
|
Price Prediction Using Machine Learning Methods on the European Market of Used Cars
Dvořáček, Petr ; Baruník, Jozef (advisor) ; Vácha, Lukáš (referee)
This master's thesis proposes accurate predictions of the prices of used cars. It builds its fundamentals on the available research and broadens the academic literature by applying several modern techniques to the European market. Us- ing machine learning models and unique data, high accuracy of predictions was obtained. The precise prediction of the residual value of a used car might benefit both the buyers and the sellers, and also reduce market inefficiencies. We are not aware of any similar work in the particular field focusing on the European market. An application programming interface (API) was exploited in order to col- lect the data. Therefore, a large set of data consisting of 221,704 used car classifieds was gathered and used in various models (MLR, PCR, LASSO, De- cision Tree, Random Forests, and ANNs). This study aims to find the most precise model for estimating the prices of used cars with the help of several performance statistics (R2, RMSE, and MAE). We support the available lit- erature as the random forest approach provided the highest accuracy when predicting the used car prices. A model using ANNs seemed to be the second best in terms of predictive performances, however, required comparably much more computing power. The effects of various attributes of used vehicles on their...
|
|
|
Determinants of Used Car Prices
Žiačik, Jan ; Baruník, Jozef (advisor) ; Kukačka, Jiří (referee)
With regard to the market share, used car market is on equal footing with market with new cars. Given its relevancy, there is an incentive to better understand its inner workings. One of the questions that can be posed, relevant especially to private individuals that want to buy or sell their car, is how the prices on used car market are determined. This research question was already focus of several previous studies, nevertheless, there are several methodological issues with them, the major being that they do not deal with model uncertainty arising from large number of possible determinants of used car prices. Therefore, the goal of this thesis is to address this shortcoming by implementing a new approach, Frequentist Model Averaging. To analyze the research question, we utilize a newly collected data set of more than 470 000 used car advertisements from several different European countries from website carvago.com. In addition to well established influence of technical attributes on the valuation of car, we also find evidence that the emission standard of car or the country that the car manufacturer is originating in, have also statistically significant effect on its valuation. Further, we also find, that there are significant regional differences in effects of different attributes. JEL...
|
|
|
Machine Learning Methods in Payment Card Fraud Detection
Sinčák, Jan ; Baruník, Jozef (advisor) ; Vácha, Lukáš (referee)
Protection of clients from fraudulent transactions is a complicated task. Banks tend to rely on rule-based systems which require manual creation of rules to identify fraud. These rules have to be set up by employees of the bank who need to look for any trends in fraudulent transactions themselves. This thesis deals with the problem of detection of fraudulent card transactions as it com- pares multiple machine learning models for fraud detection. These models can find complex relationships in the data and potentially outperform standard fraud detection systems, Logistic regression, neural network, random forest, and extreme gradient boosting (XGBoost) models are trained on a simulated dataset that closely follows properties of real card transactions. Performance of the models is measured by sensitivity, specificity, precision, AUC, and time to predict on the testing dataset. XGBoost shows the highest performance among the tested models. It is then compared to a standard fraud detection system used in a Czech bank. The bank system achieves higher specificity but XGBoost still shows promising performance. It is possible that certain machine learning models could outperform today's fraud detection systems if they are well-tuned. JEL Classification G21, K42 Keywords machine learning, card fraud, fraud...
|
guest ::
RSS feed.