Contributions in Merlin

null

The objective of this thesis is to examine viable return-generating trading strategies in the high frequency foreign exchange space using shifts in news sentiment momentum as signals. More specifically, this thesis backtests various moving average crossover strategies of hourly Euro-specific news sentiment using the TRNA dataset with over 70’000 observations. The crossover events serve as entry (exit) signals to assume long (short) positions in the EUR/USD exchange rate. In line with the findings of prior literature, an OLS regression analysis reaffirms the existence of a predictive relationship between Euro-specific news sentiment and the EUR/USD exchange rate. A backtesting exercise evaluates both simple and exponential moving average crossover strategies for three different combinations of crossover periods (2-10-hour, 2-20-hour, and 2-50-hour) with varying levels of transaction costs. The moving average crossover approach reduces the number of transactions and consequently brings down transaction costs. The results show that both sets of strategies outperform the buy-and-hold benchmark strategy and yield consistent and positive returns for the 2-20-hour and the 2-50-hour crossover periods, even after taking transaction costs into account.

Cryptographic ransomware encrypts files and demands a ransom for their decryption. Ransomware is increasingly targeting Internet of Things (IoT) devices that contain critical data. Due to limited resources, IoT devices cannot implement resource-intensive protection mechanisms to defend against ransomware. To provide a lightweight ransomware protection mechanism for IoT devices, three overlay file systems have been implemented. The overlay file systems use moving-target defense techniques to hide file type identification, increase encryption time, and trap ransomware in infinite directories. The evaluation results show that the implemented overlay file systems provide protection against ransomware attacks. The main limitations of the overlay file systems are the inability to distinguish between malicious and non-malicious applications and the performance overhead for small file sizes.

The explosive growth of the IoT has come along with an increase of cyberattacks with ransomware, rootkits and Command-and-Control malware being particularly common families. One promising approach for mitigation is offered by Moving Target Defense (MTD), which works by dynamically altering a target’s attack surface. However, the state of IoT MTD is still immature, especially lacking research dedicated to coordinating multiple MTD techniques in real applications. As a means to optimize such a system, this work explores the application of reinforcement learning (RL) to reactively deploy MTD techniques against the aforementioned malware families in a real crowdsensing scenario. First, the task of RL-based MTD selection is analyzed to distill major system requirements. Thereafter, three training simulations are presented along with the implementation of a complete, online MTD agent. As online RL is costly, the simulations gradually shift from a rather theoretical perspective towards approximating reality to allow policy transfer to a real environment. Using a supervisor to create reward signals, the first simulation marks a baseline. The second exchanges this supervisor for an anomaly detection component. For comparability both simulations use a new dataset of raw attack behaviors. The third simulation also leverages anomaly detection, yet utilizes a second dataset of behaviors monitored by a real online agent. While the agent of the first simulation learns to select MTD techniques against all attacks of the aforementioned families, the second and third simulations show that a realistic agent’s convergence is affected by anomaly detection inaccuracies, but generally attacks are effectively mitigated. Finally, implications of the online agent are discussed and its resource consumption is evaluated on a Raspberry Pi 3. Requiring less than 1MB storage and always utilizing below 80% of the available CPU and RAM, hardware poses no limitation. However, the time required to learn new attacks may impair viability.

With the rise of machine learning and artificial intelligence, new methods and resources for improving investment performance emerged and quickly gained popularity, owing to various benefits over human traders. In this thesis, two reinforcement learning agents are introduced to trade the S&P 500 E-Mini and the 10-Year US Treasury Notes futures contracts. The agents are trained and tested on 11 years of intraday data with added technical indicators. The agents’ performance is compared to a long-only benchmark. Our findings show that our agents did not outperform the benchmark and that they were unable to complete one testing session without overspending.

Recent research has placed enormous interest in the accurate prediction of stock prices. By extending conventional regression techniques, this thesis implements the concept of stress testing on an equity level using several machine learning methods, including trees, gradient boosting and neural networks (long short-term memory). The models are adapted using a novel time-series approach intended to preserve stationarity and extend the neural network using a bidirectional layer and weight penalization. Stock price changes are being forecasted and subject to severe scenarios, as commonly used on banking levels. Trees and gradient-boosted models have been identified to yield superior results in terms of common loss metrics. The implementation of macroeconomic shock scenarios reveals that predicted stock price changes are less dependent on severe scenarios using trees and gradient-boosted models; however, they are affected by material changes using neural networks. It is further investigated which macroeconomic parameters are of significant fraction in terms of forecasting stock price changes, whereas inflation, treasury and unemployment rates are among the most robust features.

This thesis examines the effect of board gender diversity on European banks’ ESG performances while considering endogeneity. The findings of the GMM on a significant positive effect highlight the importance of increasing the share of female directors in creating a sustainable banking sector. However, the positive connection is neither due to stereotypical characteristics of women nor a significant gap in values between genders of directors. Rather differences in skills and backgrounds caused by the double glass ceiling are a valid explanation. We conclude that banks should focus on overall board diversity, not just on the gender of their board members.

The effects of the Corporate Sector Purchase Programme, a quantitative easing measure introduced by the European Central Bank in 2016, is investigated in this paper. A difference in differences approach is used to estimate the effect on eligible bond yields and bid-ask spreads. A decrease in yields and a mixed effect on bid-ask spreads of eligible bonds are observed. However, these findings are not robust to alternative measures of estimation and hence cannot conclusively confirm the findings of previous academic literature.

How does the appointment of an outside CEO affect the hiring firm’s performance? Prior research reports that outside CEOs tend to underperform compared to inside CEOs, with high performance variance. Extending CEO-centric perspectives, we predict that experiential learning enhances post-succession performance, while negative transfer learning undermines it. We then offer a novel stakeholder-centric perspective, conjecturing that stakeholders’ negative sentiment toward the CEO appointment undermines post-succession performance. We further conjecture that outside CEOs are less effective in leveraging their executive experience and suffer more from negative transfer and negative sentiment compared to inside CEOs, who can leverage their familiarity and social embeddedness in the firm, which explains why outside CEOs may underperform. Analyzing the appointments of CEOs in US public firms, we find that counter to expectations, the length and breadth of their executive experience do not explain post-succession performance nor the performance differences between outside CEOs and inside CEOs. Rather, the misfit between the CEOs’ corporate background and their firms’ characteristics and the negative sentiment surrounding their appointments explain performance differences and the underperformance of outside CEOs. Accordingly, our study directs attention to the important yet previously understudied reactions of stakeholders to CEO appointments.

We examine how six search engines filter and rank information in relation to the queries on the U.S. 2020 presidential primary elections under the default—that is nonpersonalized—conditions. For that, we utilize an algorithmic auditing methodology that uses virtual agents to conduct large-scale analysis of algorithmic information curation in a controlled environment. Specifically, we look at the text search results for “us elections,” “donald trump,” “joe biden,” “bernie sanders” queries on Google, Baidu, Bing, DuckDuckGo, Yahoo, and Yandex, during the 2020 primaries. Our findings indicate substantial differences in the search results between search engines and multiple discrepancies within the results generated for different agents using the same search engine. It highlights that whether users see certain information is decided by chance due to the inherent randomization of search results. We also find that some search engines prioritize different categories of information sources with respect to specific candidates. These observations demonstrate that algorithmic curation of political information can create information inequalities between the search engine users even under nonpersonalized conditions. Such inequalities are particularly troubling considering that search results are highly trusted by the public and can shift the opinions of undecided voters as demonstrated by previous research.