Jurnal Statistika dan Komputasi

Explainable Gradient Boosting for GPA Prediction with SHAP Aggregation

2026-07-06T06:26:34+00:00

Background: The Grade Point Average (GPA) is a complicated quantity impacted by psychological factors as well as the learning environment. Previous research has concentrated exclusively on understanding GPA prediction models through a singular predictive framework, resulting in inconsistent assessments of varied significance.

Objective: This study aims to propose a SHAP aggregation framework to explain the importance of features from three gradient boosting models, consisting of XGBoost, LightGBM, and CatBoost.

Methods: The dependent variable in this study is student GPA, whereas the predictor variables encompass emotional intelligence, adversity quotient, metacognitive ability, motivation, self-efficacy, and learning environment. The predictor variable was a composite score obtained from the average of the items of each latent variable. Data were collected through a survey of 270 students with probability sampling.

Results: The SHAP Importance Feature of the three gradient boosting models exhibited varying outcomes regarding the variables with the greatest and second-highest contributions. XGBoost and LightGBM recognized metacognitive capacity as the predominant attribute, while CatBoost highlighted adversity quotient. The SHAP aggregation results indicate that Adversity Quotient is the predominant factor, with a mean importance feature value of 0.0318.

Conclusion: The findings indicate that the SHAP aggregation framework offers a more stable and robust interpretation of variable importance compared to a single model. This methodology can also serve as an analytical framework in educational data mining research.

Determinants of the Working Poor in the Papua Region of Indonesia: A Multilevel Logistic Regression Analysis

2026-07-06T06:26:42+00:00

Background: The working poor are employed individuals who fail to achieve a minimum standard of living, reflecting that employment does not guarantee economic well-being. The Papua Region of Indonesia exemplifies this paradox.

Objective: This study examines the determinants of working poverty at individual and regional levels in 2023.

Methods: A multilevel binary logistic regression is applied to hierarchical data of workers nested within regions.

Results: Approximately 23.22% of workers are classified as working poor. Male workers are 1.59 times more likely to experience working poverty than females. Rural workers face 1.57 times higher odds compared to urban workers. Informal employment substantially increases vulnerability, with workers 3.21 times more likely to be poor. Those working less than 35 hours per week have 2.06 times higher odds of poverty. Education shows a strong protective effect: workers with primary and secondary education are 2.69 and 2.10 times more likely, respectively, to be poor compared to tertiary graduates. Workers without training are also more vulnerable (OR=1.16). At the regional level, higher GRDP reduces the likelihood of working poverty (OR=0.981).

Conclusion: Reducing working poverty requires improving job quality, expanding education and training, and promoting inclusive regional economic growth.

Bivariate Inverse Gaussian Regression on Stunting and Malnutrition in Children

2026-07-06T06:26:02+00:00

Background: Research on Bivariate Inverse Gaussian Regression (BIGR) as a multivariate model is currently limited to theoretical applications. Furthermore, no studies have addressed BIGR parameter estimation using optimization methods like Broyden-Fletcher-Goldfarb-Shanno (BFGS). Meanwhile, a robust analysis is needed to identify the driving factors of child stunting and malnutrition in Central Java, given their fluctuating rates. Since stunting and malnutrition are inherently correlated, BIGR provides an appropriate joint modeling for this context.

Objective: The purpose of this study is to identify the contributing variables to stunting and malnutrition in children with BIGR.

Methods: The analysis transitions from a univariate Inverse Gaussian Regression (IGR) to a BIGR framework to effectively capture the inherent dependency between the two nutritional deficiencies. Parameter estimation was performed via Maximum Likelihood Estimation (MLE), where the non-linear log-likelihood functions were numerically optimized using the BFGS.

Results: Based on the BIGR model, the correlation coefficient between stunting and malnutrition in children was 0.48, reflecting a relatively strong correlation between the two conditions. Additionally, BIGR outperformed the independent IGR model due to its smaller AIC. The model further revealed that the number of infants with early initiation of breastfeeding is the primary driving factor for both stunting and malnutrition.

Conclusion: BIGR modeling results show that the factors causing stunting and child malnutrition in Central Java are the number of infants with early initiation of breastfeeding.

Multimodal Deep Learning for Online Gambling Promotion Detection on Indonesian Social Media

2026-07-06T06:25:46+00:00

Background: Online gambling promotion has become a major social problem in Indonesia, generating billions of rupiah in annual transactions despite strict legal prohibitions. Existing detection methods primarily focus on textual content while overlooking the visual information commonly used in social media promotions.

Objective: This study proposes a multimodal deep learning approach that combines image and text representations for detecting online gambling promotions on Indonesian social media and introduces the first publicly available multimodal dataset for this task.

Methods: A dataset of 5,028 labeled image-caption pairs was collected from Facebook, TikTok, and X, comprising 2,195 gambling promotion and 2,833 non-promotion samples. Three image embedding models, three text embedding models, and six multimodal fusion strategies were evaluated using a lightweight Multilayer Perceptron classifier.

Results: SigLIP 2 achieved the best image-only performance, while Multilingual E5-large achieved the best text-only performance. The contrastive similarity fusion strategy achieved the highest overall performance, with 97.51% accuracy and a 97.18% F1-score.

Conclusion: The proposed multimodal approach effectively detects online gambling promotions by leveraging complementary visual and textual information. The publicly available dataset also provides a benchmark to support future research.

Aggregate Loss Modeling for Renewal Gross Premium Estimation in Group Health Insurance

2026-07-06T06:25:31+00:00

Background: Group health insurance policy renewals require insurers to re-evaluate premiums based on historical claim experience. Medical inflation heightens uncertainty by driving up future healthcare costs.

Objective: This study estimates the renewal gross premium for a group health insurance portfolio using an aggregate loss risk modeling approach.

Methods: Historical claim data covered 912 insured individuals, with 102 claimants and 174 claims recorded between July 2024 and June 2025, drawn from a corporate client of an Indonesian insurance company. Claim frequency was modeled using the Zero-Inflated Poisson–Lindley (ZIPL) distribution to handle excess zeros and overdispersion, while claim severity was modeled with the Mixture Gamma–Rayleigh Distribution (MGRD) to capture medical cost heterogeneity. Parameters were estimated via Maximum Likelihood Estimation, and a 95% confidence interval for expected severity was derived using Chebyshev's inequality. The model assumes a 16.2% medical inflation rate, a 6.5% annual interest rate, and an 18% loading factor.

Results: At a 95% confidence level, the estimated renewal gross premium ranges from IDR 2.053 billion to IDR 4.352 billion. Sensitivity analysis confirms that premiums rise with increasing medical inflation.

Conclusion: The interval-based approach provides a statistically grounded numerical basis for underwriting decisions and negotiations in group health insurance renewals.

Time Series Anomaly Detection of International Tourist Arrivals Using Copula-Based Outlier Detection Method

2026-07-06T06:26:13+00:00

Background: An outlier or anomaly is an observation that deviates from normal historical patterns. An observation may appear normal individually, but can be identified as anomalous when evaluated through its dependence on other variables. Copula-based outlier detection (COPOD) accommodates multiple variables using empirical marginal distributions and tail dependence structures to identify anomalies.

Objective: This study aims to detect anomalies in the number of international tourists in Indonesia by considering the variables of inflation and the rupiah exchange rate, as well as to evaluate the handling of anomalies on the forecasting performance of long short-term memory (LSTM).

Methods: Monthly data from January 2000 to December 2025 obtained from CEIC Data, Statistics Indonesia, and Bank Indonesia were used in the study. The analysis includes data exploration and the development of feature engineering, anomaly detection using COPOD, followed by LSTM forecasting.

Results: Detection was carried out based on twelve variables resulting from feature engineering, and eleven periods were identified as anomalies. The forecasting results show better accuracy in the model after handling with a mean absolute percentage error, root mean square error, and Pearson correlation between actual and predicted data, which were 7.494%, 99233, and 0.864, respectively, on the test data.

Conclusion: That good accuracy result can’t be separated from the precision in detecting anomalies. Further research is expected to add relevant variables and develop feature engineering.

Spatial Dependence and Spillover Effects on Very Low-Income Agricultural Enterprises at the Provincial Level in Indonesia

2026-07-06T06:25:36+00:00

Background: Agricultural welfare disparities across Indonesian provinces remain substantial, as reflected in the high percentage of agricultural enterprises with very low income. These disparities are influenced by regional characteristics and spatial interactions among neighboring provinces.

Objective: This study analyzes the determinants of the percentage of agricultural enterprises with very low income at the provincial level while accounting for spatial dependence and spillover effects.

Methods: Cross-sectoral data from 38 provinces in Indonesia were obtained from the 2024 Agricultural Economic Survey. The analysis was conducted using spatial regression models, the Spatial Lag Model (SLM) and the Spatial Error Model (SEM). The spatial weight matrix was constructed using the KNN approach with k = 5. Moran's I test was used to detect spatial autocorrelation. The LM test and AIC were used to select the best model.

Results: Significant positive spatial autocorrelation was detected in the dependent variable and OLS residuals. The SLM was selected as the best model. The explanatory variables were shown to have both direct and indirect effects through spatial spillover effects.

Conclusion: Incorporating spatial dependence is essential for effective agricultural policymaking. Coordinated interprovincial policies on credit access, production facilities, and agricultural input supply are needed to improve agricultural welfare.

Stability of SHAP-Based Feature Importance Ranking under Class Imbalance, Feature Correlation, and Feature Dimensionality

2026-07-06T06:25:06+00:00

Background: Gradient boosting machine learning models, such as XGBoost and LightGBM, are widely used because of their high predictive performance. However, their complexity requires interpretation methods. SHAP is often used to explain feature contributions, but the stability of its interpretations may be affected by data characteristics and the model used.

Objective: This study evaluates the stability of SHAP-based feature importance rankings in XGBoost and LightGBM under various data conditions.

Methods: The study used controlled simulation and empirical BPJS Kesehatan claims data. In the simulation, 64 datasets were generated from combinations of minority class proportion, feature correlation, and number of features. The models were fitted repeatedly, and ranking stability was evaluated using Sequential Rank Agreement (SRA), where smaller values indicate more stable rankings.

Results: Higher feature correlation and extreme class imbalance reduced feature ranking stability. In the BPJS Kesehatan data, imbalance handling methods improved both model performance and interpretation stability. LightGBM with ADASYN produced a smaller SRA value (0.5350) than the model without imbalance handling (2.2534).

Conclusion: Feature correlation and class imbalance play an important role in determining the stability of SHAP interpretations. SMOTE and ADASYN improve predictive performance and increase the stability of feature importance rankings.

Modeling Adequacy Ratio in Life Insurance Using Markov Switching Approach

2026-07-06T06:24:30+00:00

Background: The life insurance sector often experiences nonlinear fluctuations during macroeconomic shocks and disasters. To capture these dynamics, nonlinear methods such as the Markov switching model are appropriate.

Objective: This study examines the effects of premium adequacy and investment returns on the claim payment ratio in Indonesia’s life insurance sector using a Markov switching approach.

Methods: This study analyzed 115 monthly observations of premium adequacy, investment returns, and claim payment ratios from January 2016 to July 2025, obtained from the Otoritas Jasa Keuangan (OJK). Three Markov switching models (MSM, MSV, and MSMV) were compared to identify the dominant source of regime-switching behavior. The two-regime Markov switching variance model was selected based on convergence and the lowest Akaike Information Criterion (AIC).

Results: The optimal model identified a volatile regime and a stable regime. The volatile regime had a 75.87% persistence probability with an average duration of four months, whereas the stable regime showed a 90.52% persistence probability and an average duration of 10.5 months. The probability of transitioning from volatility to stability (24.13%) exceeded the probability of shifting from stability to volatility (9.48%).

Conclusion: Indonesia’s conventional life insurance sector demonstrates resilience by maintaining stable conditions and recovering from periods of volatility.

Forecasting the Market Dynamics of the Invesco QQQ Trust Using Interpretable Econometric and Deep Learning Approaches

2026-07-06T06:24:14+00:00

Background: Financial time-series forecasting requires models that balance predictive accuracy and interpretability, while the advantage of deep learning over classical econometric approaches remains uncertain for highly liquid financial instruments.

Objective: This study compares interpretable econometric and deep learning approaches to identify the most suitable model for forecasting the closing price and trading volume of the Invesco QQQ Trust.

Methods: The study used 1,258 daily observations from 2019 to 2024 obtained from Yahoo Finance. Closing price and trading volume were treated as separate forecasting targets. Separate ARIMA(7,1,0) models were applied to each target and compared with Vanilla LSTM, Bidirectional LSTM, and Stacked LSTM. Performance was evaluated using RMSE, MAE, and MAPE.

Results: ARIMA recorded the lowest errors for both targets, with MAPE values of 0.82% for closing price and 19.96% for trading volume. Vanilla LSTM was the best-performing deep learning architecture, while Bidirectional and Stacked LSTM produced larger errors. The higher volume error indicates that trading activity was more difficult to forecast than closing-price movements.

Conclusion: Within the examined QQQ data, period, and model configurations, ARIMA provided more accurate forecasts than the evaluated LSTM architectures. Greater model complexity did not necessarily improve predictive performance.