Reducing complexity in multivariate electricity price forecasting

Purpose In short-term forecasting of day-ahead electricity prices, incorporating intraday dependencies is vital for accurate predictions. However, it quickly leads to dimensionality problems, i.e. ill-defined models with too many parameters, which require an adequate remedy. This study addresses this issue. Design/methodology/approach In an application for the German/Austrian market, this study derives variable importance scores from a random forest algorithm, feeds the identified variables into a support vector machine and compares the resulting forecasting technique to other approaches (such as dynamic factor models, penalized regressions or Bayesian shrinkage) that are commonly used to resolve dimensionality problems. Findings This study develops full importance profiles stating which hours of which past days have the highest predictive power for specific hours in the future. Using the profile information in the forecasting setup leads to very promising results compared to the alternatives. Furthermore, the importance profiles provide a possible explanation why some forecasting methods are more accurate for certain hours of the day than others. They also help to explain why simple forecast combination schemes tend to outperform the full battery of models considered in the comprehensive comparative study. Originality/value With the information contained in the variable importance scores and the results of the extensive model comparison, this study essentially provides guidelines for variable and model selection in future electricity market research.

A Bayesian Alternative to Synthetic Control for Comparative Case Studies

This paper proposes a Bayesian alternative to the synthetic control method for comparative case studies with a single or multiple treated units. We adopt a Bayesian posterior predictive approach to Rubin’s causal model, which allows researchers to make inferences about both individual and average treatment effects on treated observations based on the empirical posterior distributions of their counterfactuals. The prediction model we develop is a dynamic multilevel model with a latent factor term to correct biases induced by unit-specific time trends. It also considers heterogeneous and dynamic relationships between covariates and the outcome, thus improving precision of the causal estimates. To reduce model dependency, we adopt a Bayesian shrinkage method for model searching and factor selection. Monte Carlo exercises demonstrate that our method produces more precise causal estimates than existing approaches and achieves correct frequentist coverage rates even when sample sizes are small and rich heterogeneities are present in data. We illustrate the method with two empirical examples from political economy.

Bayestrat: Population Stratification Correction Using Bayesian Shrinkage Prior for Genetic Association Studies

In genetic association studies with common diseases, population stratification is a major source of confounding. Principle component regression (PCR) and linear mixed model (LMM) are two commonly used approaches to account for population stratification. Previous studies have shown that LMM can be interpreted as including all principle components (PCs) as random-effect covariates. However, including all PCs in LMM may inflate type I error in some scenarios due to redundancy, while including only a few pre-selected PCs in PCR may fail to fully capture the genetic diversity. Here, we propose a statistical method under the Bayesian framework, Bayestrat, that utilizes appropriate shrinkage priors to shrink the effects of non- or minimally confounded PCs and improve the identification of highly confounded ones. Simulation results show that Bayestrat consistently achieves lower type I error rates yet higher power, especially when the number of PCs included in the model is large. We also apply our method to two real datasets, the Dallas Heart Studies (DHS) and the Multi-Ethnic Study of Atherosclerosis (MESA), and demonstrate the superiority of Bayestrat over commonly used methods.

Polymorphism of FGD4 and myelosuppression in patients with esophageal squamous cell carcinoma

Background: Chemotherapy-related adverse events may restrain taxane/cisplatin administration as a regimen for patients with esophageal squamous cell carcinoma. Genetic polymorphisms may contribute to adverse event susceptibility. Method & results: The authors genotyped ten SNPs from five genes (rs1045642, rs2032582 and rs3213619 of ABCB1; rs2231137 and rs2231142 of ABCG2; rs246221 of ABCC1; rs3740066 of ABCC2; and rs10771973, rs12296975 and rs1239829 of FGD4) in 219 patients with esophageal squamous cell carcinoma treated with taxane/cisplatin. Patients with severe toxicities were compared with those with minor or no adverse events by unconditional logistic regression models and semi-Bayesian shrinkage. After adjustment for age and sex, with the null prior, FGD4 rs1239829 was statistically significantly related to grade 3–4 leukopenia (odds ratio [95% CI] in dominant model = 1.77 [1.04–3.03]). Conclusion: The minor allele of FGD4 rs1239829 was related to grade 3–4 leukopenia in patients with esophageal squamous cell carcinoma treated with taxane/cisplatin, with unclear biological mechanism.

Applications of Bayesian Shrinkage Prior Models in Clinical Research with Categorical Responses

Background: Prediction and classification algorithms are commonly used in clinical research for identifying patients susceptible to clinical conditions like diabetes, colon cancer, and Alzheimer’s disease. Developing accurate prediction and classification methods have implications for personalized medicine. Building an excellent predictive model involves selecting features that are most significantly associated with the response at hand. These features can include several biological and demographic characteristics, such as genomic biomarkers and health history. Such variable selection becomes challenging when the number of potential predictors is large. Bayesian shrinkage models have emerged as popular and flexible methods of variable selection in regression settings. The article discusses variable selection with three shrinkage priors and illustrates its application to clinical data sets such as Pima Indians Diabetes, Colon cancer, ADNI, and OASIS Alzheimer’s data sets. Methods: We present a unified Bayesian hierarchical framework that implements and compares shrinkage priors in binary and multinomial logistic regression models. The key feature is the representation of the likelihood by a Polya-Gamma data augmentation, which admits a natural integration with a family of shrinkage priors. We specifically focus on the Horseshoe, Dirichlet Laplace, and Double Pareto priors. Extensive simulation studies are conducted to assess the performances under different data dimensions and parameter settings. Measures of accuracy, AUC, brier score, L1 error, cross-entropy, ROC surface plots are used as evaluation criteria comparing the priors to frequentist methods like Lasso, Elastic-Net, and Ridge regression. Results: All three priors can be used for robust prediction with significant metrics, irrespective of their categorical response model choices. Simulation study could achieve the mean prediction accuracy of 91% (95% CI: 90.7, 91.2) and 74% (95% CI: 73.8,74.1) for logistic regression and multinomial logistic models, respectively. The model can identify significant variables for disease risk prediction and is computationally efficient. Conclusions: The models are robust enough to conduct both variable selection and future prediction because of their high shrinkage property and applicability to a broad range of classification problems.