DOI: 10.24818/jamis.2025.04005
Vol. 24, No. 4, pp. 651-672, 2025
© 2025. This work is openly licensed via CC BY 4.0.
Author(s): Alexey Litvinenko1,a, Anna Litvinenkob and Samuli Saarinenc
a University of Tartu, School of Economics and Business Administration, Estonia
b Tallinn University of Technology, School of Business and Governance, Department of Business Administration, Estonia
c Estonian Business School, Estonia
1 Corresponding author: Alexey Litvinenko, School of Economics and Business Administration, University of Tartu, Narva mnt 18, Tartu, Estonia, email addresses: alexey.litvinenko@ut.ee; alx199@gmail.com.
Keywords: R programming, econometric analysis, financial analysis, credit risk forecasting.
JEL codes: C22, G32
Abstract
Research Question: Which of the four models (MLR, IV, ARIMA, ES) performed through R programming are more precise in credit risk forecasting based on financial ratios and possess improved robustness and generalizability as well as being less prone to overfitting?
Motivation: Traditional econometric models used in credit risk forecasting often suffer from overfitting, particularly when applied to financial ratio data with low variance. This challenge is especially pronounced in small sample settings typical of emerging markets or firm-level analysis. Exploring alternative, more adaptive models is necessary to improve forecasting reliability under such constraints.
Idea: This study evaluates whether transforming financial statement data into time-series ratio formats and applying ARIMA and ES models can enhance forecasting robustness and reduce overfitting, compared to conventional linear models.
Data: The historical panel data for 7 years from the annual reports of two production companies listed on the Baltic Stock Exchange, processed into financial ratios for forecasting 3-year horizons.
Tools: All four models are developed using R programming. Forecast performance is evaluated using Akaike Information Criterion (AIC) and other diagnostic measures for predictive accuracy, robustness, and resistance to overfitting.
Findings: ARIMA and ES models demonstrate superior predictive accuracy and robustness, especially in small-sample conditions. They respond better to structural changes and recent data trends than Multiple Linear Regression (MLR) and Instrumental Variable (IV) models. This suggests ratio-based forecasting benefits from dynamic, time-sensitive modelling. The findings challenge linear assumptions and emphasize the value of time-series approaches in improving credit risk estimation under constrained data conditions.
Contribution: The study offers a replicable, R-based framework for robust credit risk forecasting, advancing time-series methods in small-sample financial analysis.
Full paper at: http://online-cig.ase.ro/jcig/art/5 Litvinenko.pdf
