Contributions in Merlin

In an ever more open and globalized world economy with unprecedented access to vast and increasing flows of information driven by the digital revolution of the 1990’s, competition has spiked between market players fighting for scarce resources. This is particularly the case of financing economic ventures. Investors require higher returns from their investments pressuring ventures in need for capital to promise spectacular earnings that are yet to be met. This situation on one hand pushes for innovation, but, at the same time, exposes the economy to greater uncertainty and risk, leading to a dangerous inflation of the financed projects’ potential. Furthermore, funding from traditional business lenders, such as banks and credit unions, usually requires a lengthy and complex vetting procedures and requirements that is complex for start-ups and young-stage ventures to meet when apply for funding. The need to address demand for funding from the growing number of start-ups and ventures has created a market for alternative forms of investments that are flourishing and aim to streamline the processes of application, approval, collateral, control, and repayment. The increased number of start-ups, characterized by huge expansion potential, spirit of innovation, high dynamism and government and university backing, has greatly incentivized a shift in financing methods. At the same time, the advent of these alternative forms of investments, characterized by less stringent requirements and controls and by a tendency to grant loans to riskier potentially more profitable ventures, has led in some cases to unsustainable promises that resulted in disastrous outcomes. In this context, we will focus on the health technology sector, that is highly promising and growing, but also subject to the frontier of controversial ethical and societal dilemmas. Although the potential benefits of health technologies are huge, they raise deep societal, ethical and moral questions and significant challenges regarding privacy rights and obligations to patients. To ground the discussion and provide a practical perspective, we will analyze the recent fiasco of Elizabeth Holmes’ Theranos. We will try to identify and understand the factors that led the 2 company, advertised as a breakthrough in health technology, to fraud and significant value loss.

The thesis uses bitcoin data to estimate parameters, with which it is then tried to assess whether the model proposed in Gabaix and Koijen (2021) can explain the mean and standard deviation of bitcoin returns. The model uses the price elasticity of demand to generate observed prices and is stated in perturbations around a baseline. The results suggest that the model does not account for the full picture of return patterns observed. The moments resulting from the simulation are far larger in absolute terms than the observed moments and the estimation results are hardly statistically signi cant. Albeit, a parameter of the model, the speed of mean reversion of flows, has not been estimated, a back-of-the-envelope calculation implies a negative and very large value to counter the effects of the estimation results. A sensible estimation of this parameter is beyond the scope of this thesis. Bitcoin is well suited for price elasticity estimations because it is only marginally influenced by supply shocks i.e., all the shocks can be ascribed to the demand side. Moreover, the full record of transactions is in theory available to the public. A wide range of studies point to the effects trading has on prices, yet only recently has literature emerged arguing for inelastic  financial markets. The model presented in Gabaix and Koijen (2021) uses flows to funds to drive asset prices away from their fundamental values. This mechanism influences the stochastic discount rate in the model. For the estimations, two datasets are used. One dubbed `daily data' is provided by Stütz et al. (2020).1 This dataset consists of bitcoin transaction data aggregated on a daily and an entity level, transactions of bitcoin holders with only one address are omitted. The second called `block-level data' has been downloaded through the GraphSense API (Haslhofer et al., 2021). This dataset has been aggregated on a block-level and contains all transactions. Both datasets entail information about on-chain transactions. With this information, the estimations are performed on the transformed data. The data used is stated in deviations from its rolling averages. The rolling averages serve as a baseline. The results of the estimations suggest a positively inelastic demand and contrarian investing agents. The simulations suggest that the model overpredicts the volatility in bitcoin markets and overstates the premium paid by investors, in absolute terms.

Semantic information of 3D scans is a fundamental step for many scene understanding applications such as autonomous driving and 3D reconstruction. The goal of this work is to obtain the semantic segmentation of indoor point clouds using deep learning techniques. While semantic segmentation with deep learning of 3D point clouds still faces several challenges. In this work, instead of directly using deep learning on 3D point clouds, we first extract the panoramic images from 3D point clouds. These panoramic images are then used as input to a CNN model for semantic segmentation. Finally, we map the obtained segmentation panoramas to the 3D point cloud and develop a viewer interface to visualize the point cloud and its corresponding panoramic images. Therefore, the segmented point cloud can be analyzed in the interface visually.

We develop the first top-down method to estimate the greenness of financial portfolios, in terms of alignment to the EU Taxonomy for sustainable activities. We also develop a method to estimate, at the same time, the portfolio exposure to climate transition risk. We provide sector-level, standardized and transparent coefficients for both estimates, based on definitions of greenness and transition risk that are applicable across countries. We analyse the portfolios of Euro Area investors in 2022, based on the confidential Securities Holdings Statistics of the European Central Bank. We find that, overall, the greenness of Euro Area investors’ portfolios is lower than their exposure to transition risk (2.8% vs. 11.7%). Across financial institutions, we estimate greenness and exposure to transition risk, respectively, at 3.2% and 12% for investment funds, at 0.8% and 5% for banks and at 4.8% and 15.1% for insurers. Our analysis also shows that investors with large amounts invested in green activities can have at the same time large exposures to transition risk.

Neural methods have shown great success in string transduction tasks such as grapheme-to-phoneme conversion or morphological inflection. A particularly successful class of models for these tasks are recurrent neural transducers that use an encoder-decoder architecture to predict character-level edit actions. This work builds on such a neural transducer. Despite its ongoing success, this particular approach offers room for improvement: The model uses an outdated software framework, and the implementation is tailored toward the use on a CPU. This leads to low training efficiency preventing the application to large datasets. Moreover, the fully recurrent structure of the model does not reflect recent technological developments, as the non-recurrent transformer architectures have become dominant in many areas of natural language processing. This thesis addresses these shortcomings by reimplementing the model using the machine learning framework PyTorch, implementing GPU-supported mini-batch training and batched greedy decoding, as well as adding support for transformer-based encoders. Experimental results on standard datasets confirm the successful reimplementation. Top rankings in the SIGMORPHON 2022 Shared Task on Morpheme Segmentation (featuring training sets of up to 750,000 samples) demonstrate the model's ability to scale: GPU training and greedy decoding are up to 250 times, respectively, 10 times faster. While transformer-based encoders rarely outperformed recurrent encoders, the initial experiments in this work lay the foundation for further experimentation.

As the Internet-of-Things (IoT) see rapid growth, in device numbers and use cases, standardization is very much lacking in the field. With heterogeneous edge devices, or so-called microcontrollers (MCUs), and a variety of operating systems (OS) to choose from, interoperability is suffering. The goal of this thesis is to port the kernel of the open-source operating system Linux onto tiny MCUs. By doing so abstracting the hardware from the application layer, and therefore providing much-needed standardization in the IoT ecosystem. This was achieved by finding the correct toolchain, and compiling the Linux and μClinux kernel with the help of tools such as Buildroot. Subsequently, the compiled distributions were tested with QEMU and ported to STM32L476G-Eval board and ESPEYE respectively. Additionally, a different approach with JuiceVM, a RISC-V virtual machine, running Linux was explored.

Cryptocurrencies are based on blockchain technology which is a continuously extending list of data records stored in individual blocks. The entire transaction history is accessible to everyone and allows analyzing the relationships between the transaction network and prices of the currency on exchange markets. In this thesis, the history of Dogecoin is described with a closer look on the most influential network changes implemented after the official launch in 2013. Further, the causal relationships of the network properties and its price dynamics is analyzed in order to obtain meaningful statistics. Time series analysis is used to describe the evolution of the network and the influence of the price or vice versa. Four different network representations over a time period of 1959 days are considered. By inducing an impulse on the network properties and the price the results show evidence for causal relationships between Dogecoin price movements and its transaction network. Different effects are observed for the different representations considered. In addition, the same analysis is performed for Bitcoin over the identical time period. Similarities between the two cryptocurrencies for the behaviour of users after a price increase is shown. However, differences in the response of the price after a stimulus on the network properties lead to different purposes of the currencies and the enthusiasm of different target groups.

This thesis comes up with a solution for e-learning in a community-based healthcare project conducted by ComBaCaL (Community Based Chronic Care) in Lesotho. Community-based healthcare is common in low- and middle-income countries due to the lack of physicians and often far distances to cover. Hence, health care is provided by CHWs (Community Health Workers) which most of the time only receive their initial training and are afterwards sent to the field. In place, physical training is expensive and requires a lot of effort and time from all stakeholders. This lack of continuous training leads to lower efficiency, safety and quality of the services provided by the CHWs. With the help of e-learning, new approaches are possible. During this thesis, e-learning requirements in the specific context of ComBaCaL are determined and multiple potential solutions are evaluated. The chosen solution is implemented and tested in a field experiment. Data is collected in the form of questionnaires and usage statistics. The participants show a high level of interest in the solution and agree that it provides great added value for them personally considering their ability to learn apart from the physical training sessions. As a result, there are suggestions for improvement within the application and challenges considering its further integration into ComBaCaL, which must be considered with the future scaling of the project.

Cryptocurrencies, which are cryptographic virtual currencies, are centralized networks, based on blockchain technology. This means, that they are not issued by a central authority. With cryptocurrencies becoming more and more important, their exchange gets more and more significant as well. There are different ways of trading cryptocurrencies, there are centralized or decentralized options. With the decentralized option of decentralized exchanges (DEX), there is no need for a centralized market maker or a central authority since decentralized finance (DeFi) platforms use automated market maker (AMM) protocols, involving no further third party. Related to that, smart contracts, which are automated transaction protocols, enable the creation of decentralized finance elements, which are for example liquidity pools. Liquidity pools are pools of cryptocurrencies or tokens that involve many different participants. When trading in markets with low liquidity, there is a potential risk of slippage. This would be the difference between the expected trading price and the price at which the trade is finally executed. Liquidity pools are therefore a possible solution for illiquid markets. Furthermore, they are also providing speed and convenience in the decentralized finance ecosystem. Liquidity pools therefore play a very important role in decentralized finance with automated market makers, especially when it comes to decentralized exchanges. The problem to be addressed in this thesis is the functionality of liquidity pools, decentralized exchanges with automated market maker protocols. Over the past few years there are different liquidity pools that have gained large market share, like Aave, Uniswap and SushiSwap. The main contribution of the thesis would be to consolidate the sources regarding the functionality of liquidity pools in detail. The research question would address how liquidity pools work in practice and what the difference among blockchain protocols are. The method of the thesis will be theoretical/conceptual. Since sources on the topic are sparse, the main contribution would be to consolidate the sources and review the functionality of liquidity pools.

As many markets are not perfectly liquid, there is a necessity of adequate integration of market liquidity risk in risk measures. We discuss different approaches how liquidity risk can be incorporated into the value at risk and the expected shortfall and highlight their benefits and shortcomings. The theoretical discussion is then supplemented by an empirical study for the Swiss stock market. By performing various backtesting methods we check for the validity and performance of the different approaches. We conclude that there is a need of liquidity considerations in daily risk measures, especially in times of market turbulence and for smallcap stocks. As the relative importance of liquidity risk decreases for longer time horizons, we infer that liquidity risk can be incorporated roughly into risk measures by choosing long enough time horizons depending on the nature of the security.