Sunday, Aug 3: 4:00 PM - 5:50 PM
0830
Topic-Contributed Paper Session
Music City Center
Room: CC-210
Applied
Yes
Main Sponsor
Business and Economic Statistics Section
Co Sponsors
Government Statistics Section
Presentations
We present a method to identify multiple changepoints in non-Gaussian (discrete or continuous) autocorrelated time series. A transformation technique is used to derive the likelihood of a non-Gaussian series with an assumed latent autocorrelated Gaussian series, enabling penalized likelihood methods to handle changepoint identification in the non-Gaussian scenarios. Here we present the results when the marginal distribution is a continuous proportion, modeled via a Beta distribution. Simulations demonstrating the efficacy of the methods will be shown and we explore an interesting dataset from the world of sports, the proportion of home runs hit by Major League Baseball batters from 1920-2023.
We consider the problem of simultaneous parameter estimation, group detection, and network recovery in the Grouped Network VAR model under the setting where the underlying network is sparse and unknown. Building on recent advances in homogeneity fusion for structured regression, we propose a fusion estimator that simultaneously recovers the true group and network structures with high probability, and derive nonasymptotic concentration bounds for the corresponding estimation error. On the implementation side, we formulate our estimation optimization problem as a mixed integer program and propose an iterative algorithm to solve it at scale. We evaluate the finite-sample performance of our estimator on synthetic data and validate its practical significance on two real datasets from macroeconomics and finance.
The logarithmic multiplicative error model (log-vMEM) has been useful in modeling and forecasting multivariate positive-valued financial time series. However, as the number of components in the vector increases, the number of parameters in the log-vMEM also increases, making their estimation computationally intensive. Our proposed approach describes regularized estimation via hierarchical lag structures for log-vMEM models with multivariate Gamma errors. The hierarchical lag structure based regularization is compared with using a non-convex penalty such as the Smoothly Clipped Absolute Deviation (SCAD) penalty, or the Minimax Convex Penalty (MCP). We apply the proposed method to model the joint dynamics of robust intraday realized volatility measures for Microsoft.
After the well-known Kalman filter (KF) and Kalman smoother (KS) algorithms, the simulation smoother emerges as the next key algorithm for operationalizing a linear state space model (SSM) in SSM-based data analysis. The KF and KS are typically used for model fitting, forecasting, interpolation of the response variable, and the estimation and extrapolation of latent components in the model. The simulation smoother further enhances data analysis by enabling the drawing of random samples from the joint distribution of the latent states, conditioned on the observed data. This presentation will demonstrate how the simulation smoother can be applied to practical problems, such as obtaining global (as opposed to pointwise) confidence bands for the latent components in an SSM (e.g., global confidence band for the latent level of a series) and deriving the sampling distribution of a function of the response variable forecasts.
Keywords
Simulation Smoother
State Space Model
Time Series
Longitudinal Data
Kalman Filter
Kalman Smoother