Analisis Komparatif Model LSTM, XGBoost, dan Hybrid LSTM-XGBoost dalam Prediksi Saldo Produk Simpanan Perbankan

Amalia, Shinta (2025) Analisis Komparatif Model LSTM, XGBoost, dan Hybrid LSTM-XGBoost dalam Prediksi Saldo Produk Simpanan Perbankan. Masters thesis, Institut Teknologi Sepuluh Nopember.

Text 6032232096-Master_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (6MB) | Request a copy

Abstract

Produk simpanan yang terdiri dari deposito, giro, dan tabungan merupakan komponen utama dalam struktur pendanaan perbankan yang berperan penting dalam pengelolaan likuiditas dan strategi penghimpunan dana. Ketidakakuratan dalam memproyeksikan saldo simpanan dapat menimbulkan risiko overfunding atau underfunding yang berdampak pada stabilitas keuangan bank. Untuk mengatasi hal tersebut, penelitian ini mengimplentasikan dan membandingkan tiga model prediktif berbasis machine learning, yaitu Long Short-Term Memory (LSTM), Extreme Gradient Boosting (XGBoost), dan model Hybrid LSTM–XGBoost dalam memprediksi saldo produk simpanan. Pembangunan model dilakukan melalui tahapan praproses data, pembentukan arsitektur model, optimasi hyperparameter menggunakan Bayesian Optimization, serta evaluasi kinerja dengan tiga metrik utama yaitu MAE, RMSE, dan MAPE. Model LSTM dibangun dengan dua lapisan LSTM berurutan yang dilengkapi dropout dan satu lapisan dense sebagai keluaran akhir, sedangkan model XGBoost dikembangkan melalui penyesuaian parameter untuk mencapai keseimbangan antara akurasi dan kestabilan hasil prediksi. Model Hybrid LSTM–XGBoost memanfaatkan keluaran lapisan dense LSTM sebagai masukan bagi XGBoost untuk menggabungkan kemampuan keduanya dalam mengenali pola temporal dan menjaga kestabilan hasil prediksi. Hasil penelitian menunjukkan bahwa seluruh model mampu memberikan prediksi saldo dengan tingkat kesalahan yang rendah. Model XGBoost unggul pada produk dengan pola saldo stabil seperti deposito dan giro, sedangkan model LSTM menunjukkan hasil lebih baik pada produk tabungan yang berfluktuasi secara periodik. Ketika kedua pendekatan tersebut digabungkan, model Hybrid LSTM–XGBoost memberikan peningkatan akurasi sebesar 17% pada deposito, 6% pada tabungan, dan 6% pada giro dibandingkan model tunggal terbaik sebelumnya. Secara keseluruhan, model Hybrid LSTM–XGBoost mampu menghasilkan sistem prediksi saldo simpanan yang lebih akurat, stabil, dan adaptif terhadap karakteristik data perbankan. Integrasi kedua pendekatan ini tidak hanya meningkatkan ketepatan hasil prediksi, tetapi juga memperkaya penerapan machine learning dalam analisis data keuangan. Hasil penelitian ini diharapkan dapat menjadi acuan bagi lembaga keuangan dalam merancang strategi perencanaan saldo, pengelolaan likuiditas, serta pengambilan keputusan yang lebih proaktif, efisien, dan berbasis data.
===================================================================================================================================
Deposit products, which consist of time deposits, current accounts, and savings accounts, are essential components of a bank’s funding structure and play a key role in liquidity management and fund mobilization. Inaccurate forecasting of deposit balances may lead to overfunding or underfunding risks, which could disrupt financial stability. To address this issue, this study implements three machine learning models, namely Long Short Term Memory (LSTM), Extreme Gradient Boosting (XGBoost), and a Hybrid LSTM XGBoost model, and compares their performance in predicting deposit balances accurately. The modeling process includes several stages, starting from data preprocessing, model architecture development, hyperparameter optimization using Bayesian Optimization, and performance evaluation using MAE, RMSE, and MAPE. The LSTM model was designed with two sequential layers equipped with dropout and a dense output layer, while the XGBoost model was tuned to achieve an optimal balance between prediction accuracy and model stability. The Hybrid LSTM XGBoost model was constructed by using the dense layer output from LSTM as input for XGBoost, combining the ability of LSTM to capture temporal patterns with the robustness of XGBoost in producing stable predictions. The results indicate that all models achieved relatively low prediction errors across all deposit products. The XGBoost model performed better on products with stable balance patterns, such as time deposits and current accounts, while the LSTM model provided more accurate results on savings accounts that exhibit periodic fluctuations. When both approaches were combined, the Hybrid LSTM XGBoost model consistently outperformed the individual models, improving accuracy by 17% for time deposits, 6% for savings, and 6% for current accounts. Overall, the Hybrid LSTM XGBoost model demonstrates higher predictive accuracy and stability, showing its potential as a reliable approach for forecasting banking deposit balances. The findings of this study are expected to support financial institutions in developing data driven, efficient, and adaptive strategies for liquidity and deposit management

Item Type:	Thesis (Masters)
Uncontrolled Keywords:	Deposito, Giro, Machine Learning, Peramalan, Tabungan ============ Current Accounts, Forecasting, Machine Learning, Saving Accounts, Time Deposits
Subjects:	Q Science > Q Science (General) > Q325.5 Machine learning. Support vector machines. Q Science > QA Mathematics > QA276 Mathematical statistics. Time-series analysis. Failure time data analysis. Survival analysis (Biometry)
Divisions:	Interdisciplinary School of Management and Technology (SIMT) > 61101-Master of Technology Management (MMT)
Depositing User:	Shinta Amalia
Date Deposited:	20 Jan 2026 04:12
Last Modified:	20 Jan 2026 04:14
URI:	http://repository.its.ac.id/id/eprint/129799

Actions (login required)

View Item