Journal of the Korean Statistical Society最新文献

In this study, we aim to forecast monthly stock returns and analyze factors influencing stock prices in the Korean stock market. To find a model that maximizes the cumulative return of the portfolio of stocks with high predicted returns, we use machine learning models such as linear models, tree-based models, neural networks, and learning to rank algorithms. We employ a novel validation metric which we call the Cumulative net Return of a Portfolio with top 10% predicted return (CRP10) for tuning hyperparameters to increase the cumulative return of the selected portfolio. CRP10 tends to provide higher cumulative returns compared to out-of-sample R-squared as a validation metric with the data that we used. Our findings indicate that Light Gradient Boosting Machine (LightGBM) and Gradient Boosted Regression Trees (GBRT) demonstrate better performance than other models when we apply a single model for the entire test period. We also take the strategy of changing the model on a yearly basis by assessing the best model annually and observed that it did not outperform the approach of using a single model such as LightGBM or GBRT for the entire period.

A major goal of survey sampling is finite population inference. In recent years, large-scale survey programs have encountered many practical challenges which include higher data collection cost, increasing non-response rate, increasing demand for disaggregated level statistics and desire for timely estimates. Data integration is a new field of research that provides a timely solution to these above-mentioned challenges by integrating data from multiple surveys. Now, it is possible to develop a framework that can efficiently combine information from several surveys to obtain more precise estimates of population parameters. In many surveys, parameters of interest are often spatial in nature, which means, the relationship between the study variable and covariates varies across all locations in the study area and this situation is referred as spatial non-stationarity. Hence, there is a need of a sampling methodology that can efficiently tackle this spatial non-stationarity problem and can be able to integrate this spatially referenced data to get more detailed information. In this study, a Geographically Weighted Spatially Integrated (GWSI) estimator of finite population total was developed by integrating data from two independent surveys using spatial information. The statistical properties of the proposed spatially integrated estimator were then evaluated empirically through a spatial simulation study. Three different spatial populations were generated having high spatial autocorrelation. The proposed spatially integrated estimator performed better than usual design-based estimator under all three populations. Furthermore, a Spatial Proportionate Bootstrap (SPB) method was developed for variance estimation of the proposed spatially integrated estimator.

Genetic evidence, especially evidence based on short tandem repeats, is of paramount importance for human identification in forensic inferences. In recent years, the identification of kinship using DNA evidence has drawn much attention in various fields. In particular, it is employed, using a criminal database, to confirm blood relations in forensics. The interpretation of the likelihood ratio when identifying an individual or a relationship depends on the allele frequencies that are used, and thus, it is crucial to obtain an accurate estimate of allele frequency. Each organization such as Supreme Prosecutors’ Office and Korean National Police Agency in Korea provides different statistical interpretations due to differing estimations of the allele frequency, which can lead to confusion in forensic identification. Therefore, it is very important to estimate allele frequency accurately, and doing so requires a certain amount of information. However, simply using a weighted average for each allele frequency may not be sufficient to determine biological independence. In this study, we propose a new statistical method for estimating allele frequency by integrating the data obtained from several organizations, and we analyze biological independence and differences in allele frequency relative to the weighted average of allele frequencies in various subgroups. Finally, our proposed method is illustrated using real data from 576 Korean individuals.

This paper introduces a novel approach for approximating misclassification probabilities in Euclidean distance classifier when the group means exhibit a bilinear structure such as in the growth curve model first proposed by Potthoff and Roy (Biometrika 51:313–326, 1964). Initially, by leveraging certain statistical relationships, we establish two general results for the improved Euclidean discriminant function in both weighted and unweighted growth curve mean structures. We derive these approximations for the expected misclassification probabilities with respect to the distribution of the improved Euclidean discriminant function. Additionally, we compare the misclassification probabilities of the improved Euclidean discriminant function, the standard Euclidean discriminant function, and the linear discriminant function. It is important to note that in cases where the mean structure is weighted, a higher number of repeated measurements yields better classification results with the improved Euclidean discriminant function and the standard Euclidean discriminant function, allowing for more information to be acquired, as opposed to the linear discriminant function, which performs well with a smaller number of repeated measurements. Furthermore, we evaluate the accuracy of the suggested approximations by Monte Carlo simulations.

In this paper, we propose a calibrated ConCave-Convex Procedure (CCCP) for variable selection in high-dimensional functional linear models. The calibrated CCCP approach for the Smoothly Clipped Absolute Deviation (SCAD) penalty is known to produce a consistent solution path with probability converging to one in linear models. We incorporate the SCAD penalty into function-on-scalar regression models and phrase them as a type of group-penalized estimation using a basis expansion approach. We then implement the calibrated CCCP method to solve the nonconvex group-penalized problem. For the tuning procedure, we use the Extended Bayesian Information Criterion (EBIC) to ensure consistency in high-dimensional settings. In simulation studies, we compare the performance of the proposed method with two existing convex-penalized estimators in terms of variable selection consistency and prediction accuracy. Lastly, we apply the method to the gene expression dataset for sparsely estimating the time-varying effects of transcription factors on the regulation of yeast cell cycle genes.

Shape, as an intrinsic concept, can be considered as a source of information in some statistical analysis contexts. For instance, one of the important topics in morphology is to study the shape changes along time. From a topological viewpoint, shape data are points on a particular manifold and so to construct a longitudinal model for treating shape variation is not as trivial as thought. Unlike using the common parametric models to do such a task, we invoke Procrustes analysis in the context of a nonparametric framework and propose a simple, yet useful, model to deal with shape changes. After conveying the problem into the nonparametric regression model, we utilize the weighted least squares method to estimates the related parameters. Also, we illustrate implementing this new model in simulation studies and analyzing two biological data sets. Our proposed model shows its superiority while compared with other counterpart models.

The regularization approach for variable selection was well developed for semiparametric accelerated failure time (AFT) models, where the response variable is right censored. In the presence of missing data, this approach needs to be tailored to different missing data mechanisms. In this paper, we propose a flexible and generally applicable missing data mechanism for AFT models, which contains both ignorable and nonignorable missing data mechanism assumptions. We propose weighted rank (WR) estimators and corresponding penalized estimators of regression parameters under this missing data mechanism. An advantage of the WR estimators and corresponding penalized estimators is that they do not require specifying a missing data model for the proposed missing data mechanism. The theoretical properties of the WR and corresponding penalized estimators are established. Comprehensive simulation studies and a real data application further demonstrate the merits of our approach.

In real applications, the correlated data are commonly encountered. To model such data, many techniques have been proposed. However, of the developed techniques, emphasis has been on the mean function estimation under correlated errors, with scant attention paid to the derivative estimation. In this paper, we propose the locally weighted least squares regression based on different difference quotients to estimate the different order derivatives under correlated errors. For the proposed estimators, we derive their asymptotic bias and variance with different covariance structure errors, which dramatically reduce the estimation variance compared with traditional methods. Furthermore, we establish their asymptotic normality for constructing confidence interval. Based on the asymptotic mean integrated squared error, we provide a data-driven tuning parameters selection criterion. Simulation studies show that the proposed method is more robust and efficient than four other popular methods. Finally, we illustrate the usefulness of the proposed method with a real data example.