Friday, July 2
C28: Bayesian Inference and Forecasting for Business Problems
Contributed Session
Friday, July 2
6:45 am - 8:00 am

Speakers

  • Feng Li (Speaker) Associate Professor, Central University of Finance and Economics
  • Bruno Levy (Speaker) Teaching Assistant, Insper
  • Anna Yanchenko (Speaker) PhD Candidate, Duke University
  • Suman Guha (Speaker) Assistant Professor, Department of Statistics, Presidency University, Kolkata
  • Margarita Grushanina (Speaker) PhD student, Vienna University of Economics and Business
  • Knut Are Aastveit (Speaker) Deputy Director of Research, Norges Bank
  • Nalan Basturk (Speaker) Associate Professor, Maastricht University
  • Hoang Nguyen (Speaker) postdoc, Örebro universitet
  • Jan Greve (Speaker) PhD Candidate, Vienna University of Economics and Business
  • Peter Knaus (Speaker) Pre-Doc Teaching Assistant, Vienna University of Economics and Business
  • Jamie Cross (Speaker) Postdoctoral Research Fellow, BI Norwegian Business School
  • Mike So (Chair) Professor, The Hong Kong University of Science and Technology

Documents

slides_5min

Subsessions

Friday, July 2
6:45 am - 6:50 am

Speaker

  • Feng Li (Speaker) Associate Professor, Central University of Finance and Economics

Description

This paper proposes a Bayesian distributed vector autoregressive (DVAR) model to the distributed system with the least square approximation method. The DVAR model properly handles the large p problem by introducing additional virtual observations. Our algorithm improves the computational efficiency for large T with the support of the Apache Spark platform. Furthermore, we consider the data streaming scenario and propose an efficient Bayesian updating scheme based on the DVAR model. The DVAR model is applied to the prediction of the ultra-long electricity load data and the stock market indexes data. Our empirical study shows the DVAR model performs well on both applications.

Friday, July 2
6:50 am - 6:55 am

Speaker

  • Suman Guha (Speaker) Assistant Professor, Department of Statistics, Presidency University, Kolkata

Description

Discrete-time spatial time series data arise routinely in meteorological and environmental studies. Inference and prediction associated with them are mostly carried out using any of the several variants of the linear state space model that are collectively called linear dynamic spatio-temporal models (LDSTMs). However, real world environmental processes are highly complex and are seldom representable by models with such simple linear structure. Hence, nonlinear dynamic spatio-temporal models (NLDSTMs) based on the idea of nonlinear observational and evolutionary equations have been proposed as an alternative. However, in that case, the caveat lies in selecting the specific form of nonlinearity from a large class of potentially appropriate nonlinear functions.

In absence of strong prior knowledge regarding the dynamics, the task become even more difficult. We address this problem by introducing the Gaussian random functional dynamic spatio-temporal model (GRFDSTM). Unlike the LDSTMs or NLDSTMs, in GRFDSTM the functions governing the observational and evolutionary equations are composed of Gaussian random functions. We study properties of GRFDSTM and demonstrate how model fitting and prediction can be carried out coherently using a Bayesian framework. We also conduct an extensive simulation study and apply our model to a real dataset. The results are highly encouraging.

Friday, July 2
6:55 am - 7:00 am

Speaker

Description

Bayesian factor models represent a very popular tool in analysis of high-dimensional datasets. The cumbersome task of determining the number of factors has in recent years been addressed in literature by employing nonparametric models for the automatic inference on the number of factors. Some latest works introduced division of the dataset into clusters, allowing the number of factors differ in different clusters, with automatic inference on both clusters and cluster-specific number of factors.

 

However, to our knowledge, factors are mostly assumed to be normally distributed. In reality, this assumption may prove to be too restrictive. Here, the automatic inference on the number of clusters and factors in the dataset is extended to a non-Gaussian case. We relax the assumption of normality by employing a Laplace prior on factors. Two types of shrinkage priors are considered: the multiplicative gamma process prior and the the cumulative shrinkage process, based on a sequence of spike-and-slab-distributions. The models are tested on the Eurozone countries inflation rates dataset.

Friday, July 2
7:00 am - 7:05 am

Speaker

Description

We propose a novel and numerically efficient quantification approach to forecast uncertainty of the real price of oil using a combination of probabilistic individual model forecasts. Our combination method extends earlier approaches that have been applied to oil price forecasting, by allowing for sequentially updating of time-varying combination weights, estimation of time-varying forecast biases and facets of miscalibration of individual forecast densities and time-varying inter-dependencies among models. To illustrate the usefulness of the method, we present an extensive set of empirical results about time-varying forecast uncertainty and risk for the real price of oil over the period 1974-2018. We show that the combination approach systematically outperforms commonly used benchmark models and combination approaches, both in terms of point and density forecasts. The dynamic patterns of the estimated individual model weights are highly time-varying, reflecting a large time variation in the relative performance of the various individual models. The combination approach has built-in diagnostic information measures about forecast inaccuracy and/or model set incompleteness, which provide clear signals of model incompleteness during three crisis periods. To highlight that our approach also can be useful for policy analysis, we present a basic analysis of profit-loss and hedging against price risk.

Documents

Paper Presentation
Friday, July 2
7:05 am - 7:10 am

Speaker

  • Nalan Basturk (Speaker) Associate Professor, Maastricht University

Description

In several scientific fields, like bioinformatics, financial and macro-economics, important theoretical and practical issues exist that involve multimodal data distributions. We propose a Bayesian approach using mixtures distributions to approximate accurately such data distributions. Shape and other features of the mixture approximations are estimated including their uncertainty. For discrete data, we introduce a novel mixture of shifted Poisson distributions with an unknown number of components, which overcomes the equidispersion restriction in the standard Poisson which accomodates a wide range of shapes such as multimodality and long tails. Our simulation-based Bayesian inference treats the density features as random variables and highest credibility regions around features are easily obtained. For discrete data we develop an adapted version of the RJMCMC method, which allows for an unknown number of components instead of the more restrictive approach of choosing a particular number of mixture components using

information criteria. Using simulated data, we show that our approach works successfully for three issues that one encounters during the estimation of mixtures: label switching; mixture complexity and prior information and mode membership versus component membership. The

proposed method is applied to three empirical data sets: The count data method yields a novel perspective of the data on DNA tandem repeats; the bimodal distribution of payment details of clients obtaining a loan from a financial institution in Spain in 1990 gives insight into the repayment ability of individual clients; and the distribution of the modes of real GDP growth data from the Penn World Tables and their evolution over time explores possible world-wide economic convergence as well as group convergence between the US and European countries. The results of our descriptive analysis may be used as input for forecasting and policy analysis.

Documents

Slides
Friday, July 2
7:10 am - 7:15 am

Speaker

Description

With the uncertain changes of the economic environment, macroeconomic downturns during recessions and crises can hardly be explained by a Gaussian structural shock.

There is evidence that the distribution of macroeconomic variables is fat-tailed and asymmetric.

In this paper, we contribute to the literature by extending the VAR models to account for a more realistic assumption of the multivariate distribution of the macroeconomic variables.

We propose a general class of Generalized Hyperbolic Skew Student's-t distribution with stochastic volatility (Skew-t.SV) VAR that allows us to take into account fat tails and asymmetry.

The Bayesian inference using a Gibbs sampler is extended to make inferences of model parameters.

We present evidence of fat tails and asymmetry for monthly macroeconomic variables.

The analysis also gives a clear message that asymmetry should be taken into account to have a better prediction during recession and crisis.

Friday, July 2
7:15 am - 7:20 am

Speaker

  • Jan Greve (Speaker) PhD Candidate, Vienna University of Economics and Business

Description

Bayesian mixture models have seen a limited degree of success in clustering applications due to issues regarding prior specification, estimation and suitable post-processing. A workflow of Bayesian cluster analysis based on the Mixture of Finite Mixtures (MFM) to support the validity of the posterior inference is presented. This takes into account recent developments in asymptotic theory and builds on the telescoping sampler, a general-purpose sampler for the Mixture of Finite Mixtures. Specifically, the proposed MFM model mitigates component misspecification through the use of a mixture of normals as component density. Crucially, the methodology explores a suitable setting of prior distributions and hyperparameters in light of the recently established theoretic posterior contraction properties. In addition, a method to facilitate elicitation of the induced prior on the partitions is also presented. These considerations combined with an appropriate hyperparameter specification on the component density together influence the informativeness of the prior as well as the mixing of the sampler. As such, the overall workflow serves as best practices known to us to avoid various pitfalls of posterior inference using Bayesian finite mixture models.

Friday, July 2
7:20 am - 7:25 am

Speaker

  • Peter Knaus (Speaker) Pre-Doc Teaching Assistant, Vienna University of Economics and Business

Description

Time-varying parameter (TVP) models are widely used in time series analysis for their ability to capture gradual changes in the effect of explanatory variables on an outcome variable of interest. The high degree of flexibility they offer can lead to overfitting when not properly regularized, which in turn results in poor out of sample predictive performance. On the other hand, approaches that are too restrictive risk not letting salient features of the data filter through. In light of these requirements, we propose a novel shrinkage process for sparse state space and TVP models. Building on the work of Cadonna et al. (2020) we leverage the desirable properties of the triple gamma prior and introduce a shrinkage process that aims to combine sufficient regularization with enough flexibility to capture salient features of the data.

Friday, July 2
7:25 am - 7:30 am

Speaker

  • Jamie Cross (Speaker) Postdoctoral Research Fellow, BI Norwegian Business School

Description

Vector autoregressions with stochastic volatility in both the conditional mean and variance are commonly used to estimate the macroeconomic effects of uncertainty shocks. Despite their popularity, intensive computational demands when estimating such models have made out-of-sample forecasting exercises impractical, particularly when working with large data sets. In this article, we propose an efficient Markov chain Monte Carlo (MCMC) algorithm for posterior and predictive inference in such models that facilitates such exercises. The key insight underlying the algorithm is that the (log-)conditional densities of the log-volatilities possess Hessian matrices that are banded. This enables us to build upon recent advances in band and sparse matrix algorithms for state space models. In a simulation exercise, we evaluate the new algorithm numerically and establish its computational and statistical efficiency over a conventional particle filter based algorithm. Using macroeconomic data for the US we find that such models generally deliver more accurate point and density forecasts over a conventional benchmark in which stochastic volatility only enters the variance of the model.

Friday, July 2
7:30 am - 7:35 am

Speaker

Description

In many fields where the main goal is to produce sequential forecasts for decision making problems, the good understanding of the contemporaneous relations among different series is crucial for the estimation of the covariance matrix. In recent years, the modified Cholesky decomposition appeared as a popular approach to covariance matrix estimation. However, its main drawback relies on the imposition of the series ordering structure. In this work, we propose a highly flexible and fast method to deal with the problem of ordering uncertainty in a dynamic fashion with the use of Dynamic Order Probabilities. We apply the proposed method in two different forecasting contexts. The first is a dynamic portfolio allocation problem, where the investor is able to learn the contemporaneous relationships among different currencies improving final decisions and economic performance. The second is a macroeconomic application, where the econometrician can adapt sequentially to new economic environments, switching the contemporaneous relations among macroeconomic variables over time.

Friday, July 2
7:35 am - 7:40 am

Speaker

Description

We present a case study and methodological developments in large-scale hierarchical dynamic modeling for personalized prediction in commerce. The context is supermarket sales, where improved forecasting of household-specific purchasing behavior informs decisions about personalized pricing and promotions on a continuing basis. This is a big data, big modeling and forecasting setting involving many thousands of customers and items, with heterogeneity of customer profiles and item categories. Models developed are fully Bayesian, interpretable and multi-scale, with hierarchical forms overlaid on the inherent structure of the retail setting and information flowing from aggregate to individual levels. Methodological innovations include extensions of Bayesian dynamic mixture models, their integration into multi-scale systems, and forecast evaluation with context-specific metrics. The use of simultaneous predictors from multiple hierarchical levels improves forecasts at the customer-item level of main interest. This is evidenced across many different households and items, indicating the utility of the modeling framework for this and other individualized forecasting applications.

Add to my calendar

Add to Google Add to Outlook (.ics)

Create your personal schedule through the official app, Whova!

Get Started