PREDICTING PERSONAL SAVINGS GOAL ATTAINMENT USING MACHINE LEARNING: A COMPARATIVE ANALYSIS OF CLASSIFICATION MODELS

Abstract

The current study evaluates the potential and efficiency of classification models using machine learning for individual savings goals predictions. The study was inspired by an evident gap within the academic literature on personal finance. While the phenomenon of macroeconomic saving behavior has been studied thoroughly, there is virtually no discussion of predictive modeling at the individual level based on machine-learned behavioral patterns. Modern systems used in retail banking and financial advisory services operate on static heuristic algorithms that do not take into consideration the diversity of individual financial profiles. The dataset analyzed is artificial, containing 32,424 samples and 20 initial features related to personal finance theory. Seven new behavioral features have been created based on initial ones using personal finance theory, such as expense-income ratio, savings ratio, and logarithm of total savings. The resulting feature space contains 27 dimensions. There is almost an equal distribution between the two classes (50.5% meet the savings target, while 49.5% fail to do so). Models which were considered for training and performance measurement increased in complexity as follows: Logistic Regression being the baseline model, followed by Decision Tree, Random Forest, XGBoost, and Long Short-Term Memory Neural Network. In terms of performance metrics, AUC-ROC was used primarily, while the other performance metrics included F1 Score, Precision, Recall, Accuracy, and Cross Validation AUC-ROC There was found to be a reliable pattern of performance gains with greater model complexity. The optimized XGBoost model proved to be the superior classifier, having achieved an AUC-ROC of 0.9999, an F1 Score of 0.9959, and an Accuracy score of 0.9958 – decreasing overall classification mistakes by a factor of over 22x (from 621, Logistic Regression). The LSTM network provided good performance with AUC-ROC of 0.9997, but fell short compared to the gradient boosting approaches. Using SHAP analysis on the optimal model, it was discovered that the two most predictive factors in saving goal achievement were total savings (log-transformed) and income per month. Importantly, demographic features such as age, gender, education, employment, and geography did not play a meaningful role in predictions. The research led to the conclusion that the application of machine learning, especially XGBoost along with suitable feature engineering, provides a feasible platform for personalized and proactive savings advice systems.