Deposit migration modelling with time series clustering methods

No Thumbnail Available

URL

Journal Title

Journal ISSN

Volume Title

Perustieteiden korkeakoulu | Master's thesis

Authors

Date

2024-05-20

Department

Major/Subject

Machine Learning, Data Science and Artificial Intelligence

Mcode

SCI3044

Degree programme

Master’s Programme in Computer, Communication and Information Sciences

Language

en

Pages

62 + 2

Series

Abstract

Deposits are an integral component of the commercial banking balance sheet, and understanding their developments is vital in managing banking risks. This thesis explores time series clustering methods using synthetic data generated entirely from public economic data, which provides more granular details of deposit flows compared to the aggregated deposit volume often used in existing studies. Economic indicators used in the study, both as underlying input for generating synthetic data and as variables for forecasting, include trend indicator output of Finland (KTKK), Euribor interest rate, Composite Indicator of Systemic Stress (CISS), and Standard and Poors' Composite Index return. Two different frameworks -- dynamic time warping (DTW) distance-based clustering, and Not Too Deep (N2D) clustering are utilized to compute the pairwise distances between time series. Clustering is done using various methods including agglomerative, K-medoids, Spectral, Gaussian mixture models, and K-means clustering. Forecasts are made using autoregressive moving average (ARMA), Bayesian vector autoregression (BVAR), and gradient boosting (XGBoost). The clustering results obtained showed that DTW + Spectral clustering performs the best in the segmentation task, followed by N2D-based clustering methods in terms of adjusted Rand index. Some of the clustering methods are also shown to be able to improve forecast performance of ARMA and XGBoost, producing lower mean absolute error (MAE) and cumulative MAE for those methods. Of the tested clustering methods, DTW + Spectral and N2D + K-Means create the most consistent improvement in forecast results and forecast biases.

Description

Supervisor

Laaksonen, Jorma

Thesis advisor

Uimonen, Anna-Maija
Kansanen, Kasper

Keywords

time series clustering, non-maturity deposits, liquidity risk, synthetic data, forecast, interest rate risk in the banking book

Other note

Citation

Permanent link to this item

https://urn.fi/URN:NBN:fi:aalto-202405263805

Collections

[dipl] Perustieteiden korkeakoulu / SCI