VARIANSI: Journal of Statistics and Its application on Teaching and Research

Forecasting Indonesian Nickel Reference Mineral Prices Using a Hybrid Intervention-Neural Network Model

Febiani Putri Ardhana — 2026-09-01

Time series data influenced by external events often exhibit abrupt changes and nonlinear patterns, making them difficult to model using a single forecasting approach. Hybrid models have emerged as an alternative by combining the strengths of linear and nonlinear models to improve forecasting accuracy. One approach that can be applied is the hybrid Intervention–Neural Network (NN) model. This study aims to obtain the best hybrid Intervention–NN model for forecasting Indonesian Nickel Reference Mineral Prices (Harga Mineral Acuan/HMA) using monthly data from January 2018 to February 2025, which were affected by the Russia–Ukraine conflict in February 2022. Modeling was conducted using intervention analysis with a step function combined with NN models consisting of 1, 2, 3, and 4 hidden neurons. The results indicate that the optimal model is the hybrid intervention ARIMA(1,1,0) step function model combined with a two-neuron NN, yielding an RMSE value of 1,074.3652 and a MAPE value of 4.7735%. Forecasts for the period from March 2025 to February 2026 indicate a downward trend in nickel prices.

Negative Binomial Regression for Low Birth Weight (LBW) Cases in Kolaka Regency

Rista Utami — 2026-09-02

Low birth weight (LBW) is a major risk factor for neonatal mortality. Kolaka Regency experienced a 41.98% increase in LBW cases in 2024, necessitating an analysis of influencing factors. This study aims to model the average number of LBW cases using negative binomial regression. Data were obtained from the Kolaka District Health Office with 12 sub-districts as observation units. The dependent variable was the number of LBW cases, while independent variables included the percentage of K4 antenatal care coverage (X₁), number of pregnant women consuming 90 iron supplement tablets (X₂), and number of pregnant women with chronic energy deficiency (CED, X₃). Results showed overdispersion in the data, making negative binomial regression appropriate. Two variables significantly influenced LBW cases: the number of pregnant women consuming 90 iron supplement tablets had a negative effect (β₂ = -0.0025; p = 0.004), and the number of pregnant women with CED had a positive effect (β₃ = 0.0231; p < 0.001). Increasing consumption of 90 iron supplement tablets reduces the average number of LBW cases, while increasing the number of pregnant women with CED increases the average number of LBW cases in Kolaka Regency.

The Effects of Poverty, Average Years of Schooling, and Expected Years of Schooling on the Net Enrolment Rate for Senior High School and Equivalent Levels in the Sulawesi and Maluku Regions, 2018–2023

Rahul Fikri Alamsyah Rahul — 2026-09-06

The Net Enrollment Rate (NER) for senior high school and equivalent levels in Indonesia still shows disparities, particularly in Maluku and Sulawesi, where rates fall below the national average, posing a challenge to national development. Poverty rates, average years of schooling, and expected years of schooling are estimated to influence the Net Enrollment Rate (NER) for senior high school and equivalent levels. The method used is panel data regression analysis because the indicators under study exhibit numerical instability and changes take time to occur. The estimation approaches used in the panel data analysis are the Common Effects Model (CEM), the Fixed Effects Model (FEM), and the Random Effects Model (REM). The best model was selected using the Chow test, the Hausman test, and the Lagrange Multiplier test. The test results indicate that the Random Effects Model is the best fit. Based on the estimation results, the percentage of the poor population has a negative and significant effect on NER, while the percentage of the poor population has a positive and significant effect on NER. The variable expected years of schooling does not have a significant partial effect at the 5% significance level. Simultaneously, the independent variables have a significant effect on NER with an adjusted R-value of 0.4105, which means that the panel regression model indicates that the independent variables account for 41.05% of the variation in the dependent variable, while the remaining 58.95% is influenced by factors other than these three variables.

Implementation of the Hierarchial Clustering Method to Predict the Number of Domestic Tourist Trips in East Java Province

Dara Agustina — 2026-09-09

Indonesian tourism represents a wealth of natural and cultural assets that play a vital role in boosting state revenue. Data from Statistics Indonesia indicates a shift in the travel patterns of domestic tourists following the COVID-19 pandemic in 2022. East Java Province emerged as the primary destination for domestic travelers within Java. The province comprises 29 Regencies and 9 Cities, each offering its own unique tourism appeal. In this study, the travel patterns of domestic tourists across these Regencies and Cities were categorized into three clusters. Subsequently, ARIMA modeling was employed to forecast the volume of domestic tourist trips in East Java for each cluster. This information provides an estimate of future domestic travel activity in the Province, thereby assisting the government in considering improvements to accessibility and tourism management, as well as the development of new tourism offerings. The ultimate goal is to maintain the quality of the tourism sector and foster a continued desire among tourists to revisit the region.

The Contribution of Digital Advertising and Social Media to Product Purchase Decisions

Maya Saftari — 2026-09-18

This study aims to determine the contribution of digital advertising and social media activities to consumers' purchase decisions in marketplaces. The survey involved 123 active marketplace users selected purposively based on their transaction experience. Data were collected using a Likert-scale questionnaire and analyzed using Pearson correlation and multiple linear regression. Social media activities had a stronger relationship (r = 0.743) than digital advertising (r = 0.620). An R² of 59.7% indicates the proportion of purchase-decision variation explained by the model, highlighting the need for marketing optimization. Digital advertising had a significant effect with a regression coefficient of 0.267 (p < 0.05), while social media activities were the most dominant factor (β = 0.580; t = 7.954). The findings show that combining digital advertising and social media activities can help increase consumers' purchase decisions in marketplaces.

ARIMA Modeling for Forecasting Indonesia’s Per Capita CO2 Emissions and Literature Study of Photobioreactor-Based Mitigation

Henny Syahrani — 2026-09-27

Increasing CO₂ emissions from fossil fuels, industrial activities, and transportation drives climate change. Although Indonesia has committed to reducing emissions by 29% unconditionally and up to 43% with international support by 2030 through its Nationally Determined Contribution, per capita emissions rise. This study forecasts Indonesia’s per capita CO₂ emissions using ARIMA based on data from 1950-2024 from Our World in Data. Data from 1950-2009 served as training and 2010-2024 as testing. The Augmented Dickey-Fuller test indicated non-stationarity at level and stationarity after first differencing. ARIMA(0,1,10) was selected based on the lowest AIC and SC, highest adjusted R², and satisfactory diagnostics. MAPE was 13.19%. Forecasts for 2025-2030 show emissions increasing from 2.95 to 3.75 tons per capita. The study evaluates IMC-PBR, which removes atmospheric through photosynthesis while generating biomass.

Application of Truncated Spline Nonparametric Regression to Modeling the Factors Affecting Economic Growth in South Sulawesi Province

Hijriah hijriah — 2026-09-30

Nonparametric regression is a statistical method that does not require a predetermined regression function and is therefore appropriate when the relationship between the response and predictor variables does not follow a specific pattern. One of the most widely used approaches is truncated spline regression, which models the relationship using segmented polynomials connected by knot points, allowing it to flexibly accommodate changes in data patterns. This study aims to model the factors affecting economic growth across regencies and municipalities in South Sulawesi Province in 2024 using truncated spline nonparametric regression. The predictor variables include the percentage of the poor population, the Open Unemployment Rate (OUR), the Labor Force Participation Rate (LFPR), and the Human Development Index (HDI). The best model was selected based on the Generalized Cross Validation (GCV) criterion. The results indicate that the optimal model was obtained with three knot points, yielding a minimum GCV value of 0.85 and a coefficient of determination of 94.32%. All four predictor variables were found to have a statistically significant effect on economic growth in South Sulawesi Province.

Geographically Weighted Regression with Adaptive Gaussian and Bisquare Kernels for Dengue Cases in South Sulawesi

Nurul Amelia — 2026-10-21

This study aims to model the factors influencing the number of Dengue Fever (DBD) cases in South Sulawesi Province and to compare the performance of Adaptive Gaussian Kernel and Adaptive Bisquare Kernel weighting functions in the Geographically Weighted Regression (GWR) model. GWR is used because it can accommodate spatial heterogeneity, where the relationships between variables differ across locations due to variations in geographic conditions, population density, and environment. Model evaluation was conducted using the Akaike Information Criterion (AIC) and the coefficient of determination (R²). The results show that the GWR model with Adaptive Bisquare Kernel weighting performs better (AIC = 307.3784; R² = 0.9263) compared to the Adaptive Gaussian Kernel (AIC = 340.7891; R² = 0.6106). Thus, the Adaptive Bisquare Kernel is more suitable for modeling DBD cases in South Sulawesi Province as it better captures local spatial variation.

Comparison of Methods Fuzzy Time Series Chen And Fuzzy Time Series Cheng On the Closing Price Forecast of PT Bank Central Asia Tbk (BBCA) Shares

Maslahatul Khaerat — 2026-10-21

Stock price forecasting is an important aspect in financial market analysis because it can support investors in making investment decisions. This study aims to forecast the closing price of PT Bank Central Asia Tbk (BBCA) shares and compare the accuracy between the Chen and Cheng Fuzzy Time Series methods. The data used are the weekly closing prices of BBCA shares for the period January 2014 to December 2024 with a total of 566 data. The forecasting process is carried out through the stages of fuzzification, formation of fuzzy logical relationships (FLR), grouping FLRs into fuzzy logical relationship groups (FLRG), and defuzzification. Accuracy evaluation is carried out using the Mean Absolute Percentage Error (MAPE). The results show that the Chen Fuzzy Time Series method produces a prediction of BBCA's closing price in the first week of January 2025 of Rp9,440.75 with a MAPE value of 6.32%, while the Cheng Fuzzy Time Series method produces a prediction of Rp9,441 with a MAPE value of 5.74%. With a smaller MAPE value, the Cheng method is proven to be more accurate than the Chen method in predicting BBCA's closing price. This shows that giving weights to fuzzy relations in Cheng's method is able to increase the accuracy of forecasting results.

Analysis of Value at Risk (VaR) on Stock Portofolios using The T-Student Copula Method

Artono — 2026-10-21

This study aims to estimate the portfolio risk of PT Bank Rakyat Indonesia (BBRI) and PT Bank Central Asia (BBCA) stocks using the t-Student Copula approach, and to compare Value at Risk (VaR) between the two assets. This method is utilized due to the characteristics of financial data, which are non-normally distributed and exhibit extreme dependence (tail dependence). The methodology integrates the GARCH(1,1) model to account for marginal volatility and the t-Student Copula to capture the dependency structure through Monte Carlo simulation with 10,000 iterations. The results indicate a significant positive correlation, with a Kendall’s Tau of 0.3242 and a Copula correlation parameter of 0.4886. Furthermore, the degree of freedom of 12.9364 confirms the presence of tail dependence. From the portfolio risk estimation, the Value at Risk (VaR) at a 95% confidence level is 1.9774% and at a 99% confidence level is 3.1301%, with Expected Shortfall (ES) values of 2.7571% and 4.0775%, respectively. By comparing the estimation results, this study identifies differences in the risk profiles of BBRI and BBCA, driven by the individual volatility of each asset. In conclusion, the t-Student Copula provides more accurate and conservative risk estimates as it is capable of anticipating simultaneous extreme loss potential, making it a highly relevant instrument for investors in portfolio risk management within the Indonesian capital market.

Application of the Hybrid ARIMA-ANN Method to Predict the Stock Price of PT Telekomunikasi Indonesia (TLKM)

Inaya Zalsabilah — 2026-10-21

Stock prices fluctuate over time, making them difficult to predict accurately. One commonly used method is the Autoregressive Integrated Moving Average (ARIMA), but it has limitations in capturing nonlinear patterns. To overcome this, an Artificial Neural Network (ANN) was used to create a hybrid ARIMA–ANN model. This study aims to determine the best model and compare the accuracy between the ARIMA model and the hybrid model in predicting the stock price of PT Telekomunikasi Indonesia (Persero) Tbk (TLKM). The data used are daily stock closing price data for the period June 2020 to June 2025. The best ARIMA model obtained is ARIMA (2,1,3), then the residuals are modeled using ANN with the (1–3–1) architecture. The results show that the ARIMA model produces a MAPE of 1.3179% and an RMSE of 60.52, while the hybrid ARIMA–ANN model produces a MAPE of 1.2941% and an RMSE of 59.01. Thus, the hybrid model is more effective in improving stock price prediction performance.

Application of the K-Nearest Neighbor (K-NN) Algorithm for Disease Classification Based on Eye Images

St. Nadira Zalzabila Zalzabila — 2026-09-30

Eye diseases are one of the health issues that can lead to a decline in visual acuity if not detected and treated at an early stage. With advances in digital image processing and machine learning technologies, eye images can be utilised as a source of information to support the automated classification of eye diseases. This study aims to apply the K-Nearest Neighbour (K-NN) algorithm to classify eye diseases based on eye images. The dataset used in this study consists of 4,217 eye images categorised into four classes: cataract, glaucoma, diabetic retinopathy, and normal. The research stages included image pre-processing, comprising conversion to greyscale images, normalisation, segmentation, and resizing. Next, texture features were extracted using the Gray Level Co-occurrence Matrix (GLCM) method to obtain features representing the characteristics of the eye images. The data were then split into training and test sets in an 80:20 ratio before classification was carried out using the K-NN algorithm. The evaluation results show that the K-NN algorithm with K = 7 achieved an accuracy rate of 81 per cent, indicating that the model is able to classify the majority of eye images accurately.