Speakers

Estimation of a Parsimonious Multiplicative Covariance Structure

Authors: Oliver Linton, Christian Hafner and Haihan Tang

Dynamic Network Perspective of Cryptocurrencies

Authors: Li Guo, Yubo Tao and Wolfgang K. Härdle

Abstract: Cryptocurrencies are becoming an attractive asset class and are the focus of recent quantitative research. The joint dynamics of the cryptocurrency market yields information on network risk. Utilizing the adaptive LASSO approach, we build a dynamic network of cryptocurrencies and model the latent communities with a dynamic stochastic blockmodel. We develop a dynamic covariate-assisted spectral clustering method to uniformly estimate the latent group membership of cryptocurrencies consistently. We show that return inter-predictability and crypto characteristics, including hashing algorithms and proof types, jointly determine the crypto market segmentation. Based on this classification result, it is natural to employ eigenvector centrality to identify a cryptocurrency’s idiosyncratic risk. n asset pricing analysis finds that a cross-sectional portfolio with a higher centrality earns a higher risk premium. Further tests confirm that centrality serves as a risk factor well and delivers valuable information content on cryptocurrency markets.

Opening

Closing

Asset Pricing using Time-Frequency Dependent Network Centrality

Authors: Jozef Barunik, Michael Ellington

Abstract: A framework is provided where time-frequency dependent network centrality links to expected excess return of financial assets. Noting that investors trade on different horizons, it is essential to understand how the connectedness of a system influences risk premium over the short-, medium- and long-term. Viewing the market as being generated by a time-varying parameter VAR model implies that a shock to the $j$-th asset is time-frequency dependent. This creates a network of time-frequency connections among all assets in the market. We propose a new measure of time-frequency dependent network centrality and apply this to all stocks listed on the S&P500. Our findings indicate that our time-frequency dependent measures significantly price assets; particularly over the longer-term.

Opening

Closing

Portfolio enhancement based on DARA stochastic dominance

Authors: Milos Kopa, Thierry Post

Abstract: Portfolio optimization based on Stochastic Dominance (SD) is theoretically appealing, for investment strategies with asymmetric risk profiles such as equity price reversal and momentum plays. Most studies in this area are based on the second-order stochastic dominance (SSD) criterion. Unfortunately, SSD has limited discriminatory power, because it requires unanimity among all global risk averters, including those with implausible attitudes towards higher-order risk. SSD optimization therefore often produces solutions which are suboptimal for all standard utility functions.To improve the power of the analysis, the paper develops a portfolio optimization method based on Decreasing Absolute Risk Aversion Stochastic Dominance (DSD). DSD is known to be more powerful than alternative dominance criteria in several related financial applications and it is generated by the most restrictive class of utility functions acceptable to most economists.The proposed optimization method improves upon the performance of Mean- Variance optimization by tens to hundreds of basis points per annum, for low to medium risk levels. The improvements critically depend on imposing Decreasing Absolute Risk Aversion instead of Global Risk Aversion or Decreasing Risk Aversion.

Opening

Closing

Using generalized estimating equations to estimate nonlinear models with spatial data

Authors: Cuicui Lu, Weining Wang, Jeffrey M. Wooldridge

Abstract: We study estimation of nonlinear models with cross sectional data using two-step generalized estimating equations (GEE) in the quasi-maximum likelihood estimation (QMLE) framework. In the interest of improving efficiency, we propose a grouping estimator to account for the potential spatial correlation in the underlying innovations for count data and binary response data. Under mild weak dependency assumptions, results on estimation consistency and asymptotic normality are provided. Monte Carlo simulations show efficiency gain of our approach in comparison of different estimation methods. Finally we apply the GEE approach to study the determinants of the inflow foreign direct investment (FDI) to China.

Cross-border Lending Network and Systemic Risk Across Global Stock Markets

Authors: Yongshi Jie and Yinggang Zhou

Abstract: This paper examines whether systemic risk of cross-border lending network can propagate to risk spillover network of global stock market. We find that the stock market risk spillover network is more stable (fragile) if the cross-border lending network becomes more diversified when the shock is small (big) enough. Moreover, the higher the ratio of cross-border lending to the GDP of the countries which borrow, the bigger impact of risk spillover on the countries which lend more. Finally, the exponential random graph model (ERGM) is used to interpret the relationship between two networks. The connections in cross-border lending network have a significantly positive on the likelihood of spillover effect relations in global stock market.

Optimal stopping and control using reinforcement learning

Authors: Vladimir Spokoiny

Abstract: TBA.

Micro Forecasting

Authors: Matus Maciak, Ostap Okhrin, Michal Pesta

Abstract: Forecasting costs is now a front burner in empirical economics. We propose an unconventional tool for stochastic prediction of future expenses based on individual (micro) developments of recorded events. Let us think of a firm, enterprise, institution, or state, which possesses knowledge about particular historical events. For each event, there are several related payments or losses spread over time. Nevertheless, the issue is that some already occurred events do not have to be necessarily reported. The aim lies in forecasting future cash flows coming from already reported, occurred but not reported, and yet not occurred events. Our methodology is illustrated on quantitative risk assessment, however, it can be applied to other areas such as startups, epidemics, war damages, advertising and commercials, credit cards, or prescription behavior of general practitioners as discussed in the paper. As a theoretical contribution, stochastic inference for marked non-homogeneous Poisson process with non homogeneous Poisson processes as marks is developed.

Topic Sentiment Asset Pricing with DNN Supervised Learning

Authors: Hitoshi Iwasaki and Ying Chen

Abstract: We develop an innovative deep neural network (DNN) supervised learning approach to extracting insightful topic sentiments from analyst reports at the sentence level and incorporating this qualitative knowledge in asset pricing and portfolio construction. The topic sentiment analysis is performed on 113,043 Japanese analyst reports and the topic sentiment asset pricing model delivers superior predictive power on stock returns with adjusted R2 increasing from 1.6% (benchmark model without sentiment) to 14.0% (in-sample) and 13.4% (out-of-sample). We find that topics reflecting the subjective opinions of analysts have greater impact than topics of objective facts and justification of the quantitative measures.

Utility Based Model Selection and Model Averaging (Including ML Models)

Authors: Robert Navratil, Jan Vecer

Abstract: This talks presents a novel approach to model selection and model averaging based on economic theory. We study a model prediction in the form of a distributional opinion about a random variable $X$. We show how to test this prediction against any alternative views. Different model opinions can be traded on a hypothetical market that trades their differences. Using utility maximization technique, we describe such a market for any general random variable $X$ and any utility function $U$. We specify the optimal behavior of agents and the total market that aggregates all available opinions and show that a correct distributional opinion realizes profit in expectation against any other opinion, giving a novel technique for model selection. Analytical solutions are available for random variables from exponential family. We determine the distribution corresponding to the aggregated view of all available opinions, giving a novel technique for model averaging.

Nuisance-parameter-free changepoint detections

Authors: Michal Pesta

Abstract: Many changepoint detection procedures rely on the estimation of nuisance parameters (like long-run variance). If a change has occurred, estimators might be biased and data-adaptive rules for the choice of tuning parameters might not work as expected. If the data is not stationary, this becomes more challenging. The aim of this paper is to present two changepoint tests, which involve neither nuisance nor tuning parameters. This is achieved by combing self-normalization and wild bootstrap. We investigate the asymptotic behavior and show the consistency of the bootstrap under the hypothesis as well as under the alternative, assuming mild conditions on the weak dependence of the time series. As a by-product, a changepoint estimator is introduced and its consistency is proved. The results are illustrated through a simulation study. The new completely data-driven tests are applied to real data examples from finance and hydrology.

Machine learning approach to chance-constrained problems: An algorithm based on the stochastic gradient descent

Authors: Martin Branda, Lukáš Adam

Abstract: We consider chance-constrained problems where the goal is to get a highly reliable solution with respect to the realizations of random elements. We focus on the problems with a large (huge) number of scenarios. We propose a novel method based on the stochastic gradient descent which performs updates of the decision variable based only on looking at a few scenarios. We modify it to handle the non-separable objective. A complexity analysis and a comparison with the standard (batch) gradient descent method is provided. We give three examples with non-convex data which include optimal design of gas network with random demand and stochastic optimal control of electrostatic separator. We show that our method provides a good solution fast even when the number of scenarios is large.

Resampling methods in change point analysis

Authors: Zuzana Prášková

Abstract: Critical values of change point tests in linear/panel data models are usually based on limit distribution of the respective test statistics under the null hypothesis. However, the limit distribution is very often a functional of a Gaussian process and depends on unknown quantities that cannot be easily estimated. In many situations, convergence to the asymptotic distribution is rather slow and the asymptotic critical values are not well applicable in small and moderate size samples. It has appeared that resampling methods provide reasonable approximations for critical values of test statistics for detection changes in linear models. Dependent wild bootstrap is a resampling procedure for dependent data that has been developed as an alternative to existing block-bootstrap methods with the aim to mimic the dependency structure of the analyzed data not only in the blocks but in the whole sample. A variant of this method is proposed to approximate critical values of test statistics for detection a change in a dynamic panel data model with cross-sectional dependence.

RatingBot: A Text Mining Based Rating Approach

Authors: Diana Hristova

Abstract: Credit risk is at the core of banking business and its adequate measurement is crucial for financial institutions. Due to lack of historical default data and heterogeneity of customers, qualitative expert-based information is an important factor in measuring the creditworthiness of large companies. However, such information is often extracted manually, causing inefficiencies and possible subjectivity. The RatingBot is a text mining based rating approach, which efficiently and objectively models relevant qualitative information based on annual reports. It combines both the literature on text mining in finance and machine learning in c redit rating to derive the credit rating of a company. The approach is evaluated on two datasets: a publicly available one that facilitates replicability, and a dataset provided by a major European bank representing real-world scenario. The results show that RatingBot delivers additional predictive power and should be considered in future research on credit rating models.

Rating Ordinality in ML Based Credit Models

Authors: Steffen Thesdorf

Abstract: The credit rating as a measure of a company's ability to pay back its liabilities is a crucial information source, e.g. for bond investors. There have been numerous attempts to reproduce credit ratings by objective models, that is without any (presumably costly) expert judgment. A comprehensive literature review of studies on corporate credit rating models reveals that a wide range of flexible machine learning models are applied to the credit rating domain which however usually neglect the rating ordinality. The objective of present study is to find an answer to the question whether the consideration of the order information improves the prediction performance of the rating models. The Ordinal Pairwise Partitioning (OPP) approach is shown to be a straightforward enhancement to allow a nominal classification model to account for output ordinality. The nominal classification versions of the machine learning models are hence compared with their OPP counterparts by applying all models to a Deutsche Bank dataset. The results are ambiguous. Whereas Ordered Partitions (OP), as one variant of OPP, proves to promote the prediction accuracy, the OPP variants One-vs-Followers (OVF) and One-vs-Next (OVN) yield inferior outcomes.

Testing for neglected nonlinearity in the conditional quantile using neural networks

Authors: Georg Keilbar, Weining Wang

Abstract: Neural networks have proved to achieve tremendous success in a large number of prediction problems. Unfortunately, this success comes at the expense of imposing challenges for the interpretation of models and for conducting inference in general. In this paper, we propose a test for neglected nonlinearity in the conditional quantile by comparing the fit of a linear model to an alternative model constructed as a neural network series estimator. The problem can be formulated as a significance test of the output weight parameters of the neural network in the presence of nuisance parameters under the alternative. We provide consistency and asymptotic normality of the neural network series estimator. We propose three test statistics as functions of a Wald process depending on the nuisance parameters the average, the exponential average and the supremum Wald statistic. Critical values can be simulated. Simulations show that our test has proper size and good power. An application to analysis of systemic risk in a financial network indicates an important role of neglected nonlinearity.

Heterogeneous Treatment Effects through Tenure on Job Satisfaction: A Machine Learning Evaluation.

Authors: Daniel Jacob

Abstract: Scientists who give lectures or seminars might be more prone to developing new ideas for their research through the interaction with students during their teaching activities. In this paper, I estimate heterogeneous treatment effects of teaching on the perceived degree of intellectual challenge of scientists using data from the National Science Foundation (NSF). To provide as much information about the causal effects as possible and also to control for a potential selection-bias, I expand the idea of making inference on key features of heterogeneous effects sorted by impact groups (GATES), to non-randomised experiments. This is achieved through the implementation of a doubly-robust estimator. Cross-splitting with multiple iter- ations is a further extension to avoid biases introduced through sample splitting. The advantage of the proposed method is the robust estimation of heterogeneous treatment effects which is comparable with other models and thus keeps its flexibility in the choice of machine learning methods and at the same time its ability to deliver interpretable results. The empirical findings support the hypothesis that teaching activities, conducted by scientists, have a causal effect on the satisfaction with the job’s intellectual challenge. The difference in the satisfaction (scaled from 1 for very dissatisfied to 4 for very satisfied) is 0.14 between the most and least affected scientists. I conduct a classification analysis to provide insight into the average values of such key features, again for the most and least affected. Finally, I estimate the conditional average treatment effect for each individual, which allows making predic- tions of the treatment effect for new scientists. I find that there is heterogeneity in the treatment effect which is positive for the majority and negative for some scientists.

Dynamic density forecasting using machine learning

Authors: Lubos Hanus and Jozef Barunik

Abstract: The use of machine learning techniques is proposed to describe and forecast the conditional probability distribution of asset returns. We redefine the problem of forecasting of conditional probabilities looking from a different perspective than traditional ordered binary choice models. Using deep learning methods, we offer a better description of asset returns distribution. The study on the most liquid U.S. stocks shows that predictive performance of machine learning methods is promising out-of-sample. We provide a comparison of machine learning methods to the unordered and order binary choice models used by the literature.

Asset Pricing with Quantile Machine Learning

Authors: Martin Hronec and Jozef Barunik

Abstract: A large scale empirical test is performed for an asset pricing model based on agents with quantile utility preferences instead of the standard expected utility. Using machine learning methods, we predict quantiles of individual stock returns obtaining the whole forecasted distributions. We document heterogeneity in models parameters across different quantiles. We show that forecasting all quantiles together, using multi-task deep learning is better than forecasting quantiles individually. The forecasting models allow us to construct portfolios based on the whole distribution instead of just a conditional mean. We show the economic value added of looking at the whole forecasted distribution by forming quantile-based long-short portfolios, as well as favourably forecasting value-at-risk.

Uniform Inference in High-dimensional Gaussian Graphical Models

Authors: Sven Klaassen, Jannis Kueck, Martin Spindler and Victor Chernozhukov

Abstract: Graphical models have become a very popular tool for representing dependencies within a large set of variables and are key for representing causal structures. We provide results for uniform inference on high-dimensional graphical models with the number of target parameters d being possible much larger than sample size. This is in particular important when certain features or structures of a causal model should be recovered. Our results highlight how in high- dimensional settings graphical models can be estimated and recovered with modern machine learning methods in complex data sets. To con- struct simultaneous confidence regions on many target parameters, sufficiently fast estimation rates of the nuisance functions are crucial. In this context, we establish uniform estimation rates and sparsity guarantees of the square-root estimator in a random design under approximate sparsity conditions that might be of independent inter- est for related problems in high-dimensions. We also demonstrate in a comprehensive simulation study that our procedure has good small sample properties.

Cryptocurrencies

Authors: Jovanka Lili Matic

Abstract: The market for cryptocurrencies is a very dynamic market with high volatility movements and discontinuities from large jumps. We investigate the risk-management perspective when selling securities written on cryptocurrencies. To this day, options written on cryptocurrencies are not officially exchange-traded. This study mimics the dynamics of cryptocurrency markets in a simulation study. We assume that the asset follows the stochastic volatility with correlated jumps model and price options with parameters calibrated on the CRIX, a cryptocurrency index that serves as a representative of market movements. We investigate on risk-management opportunities of hedging options written on cryptocurrencies and evaluated the hedge performance under misspecified models. The hedge models are misspecified in the manner that they include fewer sources of randomness. We hedge with the industry-standard Black-Scholes option pricing model, the Heston Stochastic volatility model, and the Merton jump-diffusion model. We perform delta-hedging. We report poor hedges when calibration is poor. The results show a good performance in hedging with the Black-Scholes model and the Heston model. The study outlines the poor performance of the Merton model. We observe large unhedgeable losses in the left tail and consider these losses to result from large jumps.

Exploring the dynamic relationships between Fintech and other Traditional Finance

Authors: Yuxuan Chen, Chungchieh Cheng, Huimin Chung

Abstract: We study the dynamic relationship between Fintech and other traditional finance. During 2018, we find that the total spillover is strong, which reflects the external economic activity and financial event effects with time. In the study period, we also find the FINX is getting interconnect to other traditional finance. The FINX is not a driving force for other financial ETF spillovers, but the impact of FINX on other financial ETFs is getting stronger.

Shape-based domain adaptation via optimal transportation

Authors: Alexandra Suvorikova

Abstract: Domain adaptation problem aims at learning a well performing model, trained on a source data 𝑆 (images, vectors, e.t.c), applied then to different (but related) target sample 𝑇. Aside from being attractive due to obvious practical utility, the setting is challenging from theoretical point of view. In this work we introduce a novel approach to supervised domain adaptation consisting in a class-dependent fitting based on ideas from optimal transportation (OT) theory which considers 𝑆 and 𝑇 as two mixtures of distributions. A parametrized OT distance is used as a fidelity measure between 𝑆 and 𝑇, providing a toolbox for modelling of possibly independent perturbations of mixture components. The method is than used for describing the adaptation of immune system in humans after moving to another climatic zone.

Optimal Interest Rate Decision for a Consumer Loan

Authors: Tomas Rusy

Abstract: In this study, we present a stochastic programming asset-liability management model which deals with decision-dependent randomness. The model is designed to answer a question that is frequently asked by many (multinational) companies - what is the 'best' interest rate that a company should offer to a customer for a loan and how should it be financed? In determining the answer, we must consider numerous factors. These include the possibility of the customer rejecting the loan, the possibility of the customer defaulting on the loan and the possibility of prepayment. Moreover, these random effects have a clear relationship with the offered interest rate of the loan, which induces decision-dependent randomness. Another important factor, which plays a major role for liabilities, is the price of money at the market. This is determined by the corresponding market interest rate and we capture its evolution in the form of a scenario tree. In the presentation, we will formulate a non-linear, multi-stage, decision-dependent randomness stochastic program which will span over the lifetime of a typical consumer loan. Its solution will show us the optimal decisions the company should make. In addition, we will also show a sensitivity analysis demonstrating the results of the model for various parameter settings describing different types of customer. Finally, we will discuss the losses caused in the event that companies do not act in the optimal way.

Portfolio optimization under robust stochastic dominance constraints

Authors: Karel Kozmik

Abstract: We use modern approach of stochastic dominance in portfolio optimization, where we want the portfolio to dominate a benchmark. Since the distribution of returns is often just estimated from data, we look for the worst distribution that differs from empirical distribution at maximum by a predefined value. First, we define in what sense the distribution is the worst for the first and second order stochastic dominance. For the second order stochastic dominance, we use two different formulations for the worst case. We derive the robust stochastic dominance test for all the mentioned approaches and find the worst case distribution as the optimal solution of a non-linear maximization problem. Then we derive programs to maximize an objective function over the weights of the portfolio with robust stochastic dominance in constraints. We consider robustness either in returns or in probabilities for both the first and the second order stochastic dominance. To the best of our knowledge nobody was able to derive such program before. We apply all the derived optimization programs to real life data, specifically to returns of assets captured by Dow Jones Industrial Average, and we analyze the problems in detail using optimal solutions of the optimization programs with multiple setups. The portfolios calculated using robustness in returns turned out to outperform the classical approach without robustness in an out-of-sample analysis.

Application of Machine Learning methods in ranking and risk management

Authors: Vaclav Kozmik

Abstract: In Southeast Asian market, 90% of e-commerce volume is driven by customers who shop online for the first time and consequently traditional recommendation engine approach is not applicable. While the number of items to be offered is enormous, it is essential to have good ranking solution to present the best goods on first page of results. We will present TaRank, ranking engine based on gradient boosting which is designed to solve these issues and it has be well proven by practice – increase of click through rates by 20%. Moreover, we will focus on how to apply gradient boosting algorithm in a conservative field of risk management. We will show that stable model which significantly outperforms logistic regression can be built and that such approach is applicable not only in banking, but also in other sectors.

Random Forest Estimation of the Ordered Choice Model

Authors: Michael Lechner and Gabriel Okasa

Abstract: In econometrics so-called ordered choice models are popular when interest is in the estimation of the probabilities of particular values of categorical outcome variables with an inherent ordering, conditional on covariates. In this paper we develop a new machine learning estimator based on the random forest algorithm for such models without imposing any distributional assumptions. The proposed Ordered Forest estimator provides a flexible estimation method of the conditional choice probabilities that can naturally deal with nonlinearities in the data, while taking the ordering information explicitly into account. In addition to common machine learning estimators, it enables the estimation of marginal effects as well as conducting inference thereof and thus providing the same output as classical econometric estimators based on ordered logit or probit models. An extensive simulation study examines the finite sample properties of the Ordered Forest and reveals its good predictive performance, particularly in settings with multicollinearity among the predictors and non-linear functional forms. An empirical application further illustrates the estimation of the marginal effects and their standard errors and demonstrates the advantages of the flexible estimation compared to a parametric benchmark model.

Forecasting high-frequency stock market returns using embedded limit order book data

Authors: Marius Sterling, Niels Wesselhöfft

Abstract: In this paper we show that limit order book information for NASDAQ stocks can be utilized to forecast stock market returns in high sampling frequencies. Embedding the data of the limit order book is crucial for the model performance. Starting from linear regression with parametric, expert-judged embedding to Temporal Convolutional Nets (TCN) with raw data, we show that there is a trade-off between complexity in embedding and the complexity of the model. In accordance with the literature, using Layer Relevance Propagation (LRP) and Deep Taylor Decomposition, we show that order book levels closer the mid price are more relevant for the model performance.