Authors: Mykola Babiak and Jozef Barunik
Abstract: This paper identifies new currency risk stemming from a network of idiosyncratic option-based currency volatilities and shows how such network risk is priced in the cross-section of currency returns. A portfolio that buys net-receivers and sells net-transmitters of short-term linkages between currency volatilities generates a significant Sharpe ratio. The network strategy formed on causal connections is uncorrelated with popular benchmarks and generates a significant alpha, while network returns formed on aggregate connections, which are driven by a strong correlation component, are partially subsumed by standard factors. Long-term linkages are priced less, indicating a downward-sloping term structure of network risk.
Jozef Barunik Associate Professor, Academy of Sciences and Charles University
Jozef Baruník is an Associate Professor at the Institute of Economic Studies, Charles University in Prague. He also serves as a head of the Econometrics department at the Czech Academy of Sciences. In his research, he develops mathematical models for understanding financial problems (such as measuring and managing financial risk), develops statistical methods and analyzes financial data. Especially, he is interested in asset pricing, high-frequency data, financial econometrics, machine learning, high-dimensional financial data sets (big data), and frequency domain econometrics (cyclical properties and behavior of economic variables).
Authors: Jianqing Fan, Tracy Ke, Yuan Liao, and Andreas Neuhierl
Abstract: We develop new financial economics theory guided structural nonparametric methods for estimating conditional asset pricing models using deep neural networks, by employing time-varying conditional information on alphas and betas carried by firm-specific characteristics. Contrary to many applications of neural networks in economics, we can open the black box of machine learning predictions by incorporating financial economics theory into the learning, and provide an economic interpretation of the successful predictions obtained from neural networks, by decomposing the neural predictors as risk-related and mispricing components. Our estimation method starts with period-by-period cross-sectional deep learning, followed by local PCAs to capture time-varying features such as latent factors of the model. We formally establish the asymptotic theory of the structural deep-learning estimators, which apply to both in-sample fit and out-of-sample predictions. We also illustrate the double-descent-risk phenomena associated with over-parametrized predictions, which justifies the use of over-fitting machine learning methods.
Authors: Cathy Yi-Hsuan Chen, Jozef Barunik and Ankush Agarwal
Abstract: We consider a general framework that allows the coexistence of a stochastic volatility factor and a stochastic uncertainty factor. Motivated by the stylised facts of the volatility and uncertainty factor including mean reversion, leverage effect, and clustering, we characterise these two processes as the CIR-type process and model their interplay with the (log) price process. We introduce a novel ``bottom-up'' uncertainty measure based on cross-sectional dispersion of firm-level news-based sentiment score. We also consider a ``top-down'' measure, the economic policy uncertainty (EPU) of Baker et al (2016). Compared to the volatility process, the uncertainty process exhibits a prompt mean reversion and a modest feedback effect on the price process. We obtain a positive feedback effect from both volatility and uncertainty factors. It is worth noting that like the volatility factor the uncertainty factor exhibits a ``continuous-time'' leverage effect albeit less prominent. Concerning the correlated Brownian motions in the system, we derive the desired propositions for the volatility feedback w.r.t the two risk factors and correct the potential biases in the population regression. Using the NASDAQ news for the construction of an uncertainty measure, we calibrate the model parameters and conclude that given a salient correlated Brownian motion between the price and volatility process, the population feedback coefficient corresponding to the volatility factor is more significant than the uncertainty factor. The results are robust for the considered uncertainty measures.
Authors: Grzegorz Dudek, Slawek Smyl, Pawel Pelka
Abstract: Forecasting time series with multiple seasonal periods is a challenging problem due to the complicated relationship between input and output data. To solve this problem, a forecasting model should be equipped with appropriate mechanisms to deal with short- and long-term dynamics as well as a trend and variable variance. In this study, we show an evolution of the hybrid and hierarchical models we have developed recently for forecasting time series with complex seasonal pattern. The models are derived from the winning submission to the M4 forecasting competition and produce point forecasts as well as predictive intervals. They combine exponential smoothing (ES) and gated recurrent neural networks (RNN) in a hierarchical architecture composed of a global part trained across many time series (cross-learning) and a time series specific part. ES extracts dynamically the main components of each individual series and enables the model to learn their representation. A multi-layer RNN is equipped with dilated recurrent cells designed to efficiently model short-term, long-term and seasonal dependencies in time series. The most advanced models are based on a new attentive dilated recurrent cell, which implements an attention mechanism for dynamic weighting of input vector components. To improve the internal series representation, RNN learns simultaneously both the ES parameters and the main mapping function transforming inputs into forecasts. Our most recent solution is a contextually enhanced model which extends per-series input data with information contained in the context series (representative series) and thus supports forecasting individual series. We evaluate the models on the hourly electricity demand forecasting for 35 European countries. The results demonstrate that our models outperform both statistical and state-of-the-art machine learning models in terms of accuracy.
Authors: Dan Gabriel Anghel and Petre Caraiani
Abstract: Monetary policy shocks are known to affect financial markets. However, it is less clear how a monetary policy shock can affect their network structure. We estimate daily total connectedness and net connectedness for 10 industry portfolio indices based on intraday data. Using event-based regressions, we show that total connectedness is positively influenced by surprise changes in the interest rates. Net connectedness of some industry indices is also influenced, some in a positive, while others in a negative direction, revealing how monetary policy shocks propagate through the stock market network at a high-frequency level. The results point to the sectoral differences in the propagation of the monetary policy shocks.
Authors: Petra Laketa
Petra Laketa Researcher, Charles University
Authors: Jan Vecer
Abstract: Utility maximization depends on the choice of the underlying riskless asset as a numeraire. We show that the only numeraire invariant utility is a logarithmic function. We also note that the prices can be expressed as the likelihood ratio of the respective state price densities. On the other side, each state price density generates an asset that corresponds to the log utility optimal portfolio with respect to all assets. We show that the expected log return of the price is a relative entropy between the state price densities. When the market agent that maximizes log utility uses a mixture distribution of the state price densities of the market assets, the resulting optimal portfolio is static. When the market agent uses a prior distribution for her market opinion distribution, the resulting wealth of each parameter updates in a Bayesian fashion.
Authors: Nagy S.
Stano Nagy Professor, Charles University
Authors: Rather Miftachov
Abstract: Originating from cooperative game theory, Shapley values have become one of the most widely used measures for variable importance in applied Machine Learning. However, the statistical understanding of Shapley values is still limited. In this paper we take a nonparametric (or smoothing) perspective by proposing Shapley curves as a local measure for variable importance. We propose two estimation strategies and derive the consistency and asymptotic normality both under independence and dependence among the features. This allows us to construct confidence intervals in order to conduct inference on the estimated Shapley curves. In an empirical application, we analyze the attributes which drive the prices of cars.
Ratmir Miftachov PhD student, Humboldt-Universität zu Berlin
Authors: Martin Hronec
Abstract: We study out-of-sample performance of portfolio selecting investor with $ au$-quantile preferences. Investor's risk aversion is captured by $ au$, where more risk-averse investor maximizes lower $ au$-quantile. Using number of empirical and simulated datasets we document differences in optimal portfolios across different levels of risk aversion. We compare optimal quantile portfolios with equal-weighting and global minimum variance portfolios.
Authors: Matej Nevrla
Abstract: We propose a new model of asset returns with common factors that shift rele- vant parts of the stock return distributions. We show that shocks to such non-linear common movements in the panel of firm’s idiosyncratic quantiles are priced in the cross-section of the US stock returns. Such risk premium is not subsumed by the common volatility, tail beta, downside beta, as well as other popular risk factors. Stocks with high loadings on past quantile risk in the left tail earn up to an annual five-factor alpha 7.4% higher than stocks with low tail risk loadings. Further, we show that quantile factors have predictive power for aggregate market returns.
Authors: Jiawen Liang
Abstract: We develop a novel framework to learn investors' risk preferences with low-resolution data. Low resolution arises from a short trajectory that records investors' portfolio choices, especially in the presence of disasters. More importantly, the observed portfolio choice is contaminated by behavior bias. We start with an inverse optimization to minimise biases and infer risk aversion. We propose an algorithm that incorporates the smoothed bootstrap to inverse optimization. The smoothed bootstrap helps oversample the required data from the kernel distribution, so as to support the inference of risk aversion under disasters. Additionally, we regularize the dimension of the estimated risk aversion by the Lasso shrinkage technique. We show that the algorithm's risk aversion asymptotically converges to the true risk aversion. Based on the learned risk aversion, we adopt reinforcement learning to find the optimal investment strategies. Finally, using real data, we demonstrate that Robo-advisors adopting our approach best act on behalf of investors and mitigate the potential behavior bias.
Authors: Jana Junova
Jana Junova PhD student, Charles University
Authors: Karel Kozmik
Karel Kozmik PhD student, Charles University