Work Package N.3 deals with studying large cross-sectional samples, asset commonalities and the systemic impact with high-frequency data. WP3 will propose original methodological solutions to explore the informative content of high frequency data for uncovering commonalities and for systemic risk analysis in frameworks where the cross-sectional dimension might be large.
 
WP3 is mainly composed of two research lines.
1) The fisrt line focuses on the generalization of the tests for quantile causality proposed by Hong et al. (2009) and by Jeong et al. (2012) in a framework where high-frequency data might cause a lower frequency variable (in an econometric sense). This has implications for market monitoring and to anticipate the impact on daily market measures of systemic events (systemic co-jumps as in Caporin, Kolokolov and Renò, 2017), flash crashes, and relevant market-wide news. The main advantage of the approach we propose is its flexibility, allowing the derivation of causality tests even under partial information availability, and/or temporal misalignment between the high frequency information and the low frequency target variables.
Our approach is related to the literature on MIDAS models, see Ghysels et al. (2007), among many others, but with a specific focus on the informative content of high frequency data. From the empirical viewpoint, we plan to apply the test we will develop in related frameworks where partial daily information over a given month might be used to anticipate the effect on monthly indicators. Consequently, if high-frequency data are aggregated at the daily frequency, we propose to analyse their informative content in causing monthly systemic risk indicators, or to evaluate the impact of high-frequency-related financial information on monthly indicators of real economic activity. Such an analysis is related to the recent literature on combining systemic risk measures to identifying the most relevant ones, see Giglio et al. (2016) and Nucera et al. (2016), and to verify if systemic risk indexes cause economic variables.
2) The second research line deals with the analysis of daily measures derived from HF-data in a framework with large cross-sectional and temporal dimensions. We will extend dynamic models for realized measures (of volatility, but also of liquidity, as well as for jumps, multiple and systemic co-jumps), accounting for interdependence among variables with HAR-like structures (Corsi, 2009). Given the presence of large cross-sectional dimensions, we will follow three main approaches.In the first approach, we introduce latent factors, inspired by the works of Breitung and Eickmeier (2015), with the joint presence of global and local factors, the latter that might be active over different set of variables (local in space) or over different time zones (local in time). These latent factors will provide a view on the commonalities existent in the cross-section. Furthermore, the availability of large cross-sections will open the door for analysing the interdependence using panel methods as in Patton and Sheppard (2015). By applying these novel methodological tools on currency data, we will reconsider the exchange rate literature that points at the identification of the causes of volatility clustering and associated with the heat waves and meteor shower hypotheses of Engle et al. (1990). In the second approach, we will analyse the different degree of dynamic interdependence between realized variance measures due to the continuous and discontinuous components of the realized volatility while accounting at the same time for systematic/systemic common drivers (realized measures at the market level and/or systemic co-jumps). From the methodological viewpoint, the latter approach requires the introduction of penalization terms in the estimation step, due to the large cross-sectional dimension, with different penalty coefficients applied on different information drivers (continuous and discontinuous volatility components). We plan to extend the approach of Rothman et al. (2010) to large dimensional Bayesian sparse VAR models where different group of coefficients have their own penalty parameter (see Xu and Ghosh, 2015).
From a purely empirical perspective, the results of the models will be read with a systemic risk viewpoint. In fact, the estimated model will allow recovering financial networks monitoring the interdependence within realized measures as well as between realized measures and discontinuous jumps. Building on the spatial statistics and econometrics literature, we plan to analyse the distance between these networks by using traditional and nontraditional methods to detect spatial concentration, starting from the contributions of Conley and Topa (2002), Elhorst (2010) and Corrado and Fingleton (2012). Finally, in the third approach, we will focus on large size realized covariance matrices. The recent literature includes several proposals for measuring realized covariances among a large number of assets, see Kim et al. (2018) and therein cited references. However, very few papers tackle the issue of modelling and forecasting these matrices.
 
We plan to contribute to this recent strand of the literature pointing at approaches aimed at uncovering commonalities in realized covariances and to link these commonalities to traditional risk factors. WP3 will be valuable for market monitoring, stress test scenarios and risk management, and of course useful, given its nature, for the other work packages of HiDEA.