A machine learning model for automated forecasting of large-value payment system transactions
Payment systems are key components of the economy and their safe and efficient operation is one of the primary concern of financial system overseers. Monitoring the operations of a critical infrastructure, such as a large-value payment system (LVPS), is a key oversight activity. The observation of t...
Saved in:
Main Author: | |
---|---|
Format: | text |
Language: | English |
Published: |
Animo Repository
2015
|
Subjects: | |
Online Access: | https://animorepository.dlsu.edu.ph/etd_masteral/5084 |
Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
Institution: | De La Salle University |
Language: | English |
Summary: | Payment systems are key components of the economy and their safe and efficient operation is one of the primary concern of financial system overseers. Monitoring the operations of a critical infrastructure, such as a large-value payment system (LVPS), is a key oversight activity. The observation of transactions can be complemented by forecasts that can be used to support decision-making such as determining actions to take on a specific institution or choosing which system-wide policies to implement. The transaction of an LVPS can be sequentially measured over time and represented as a time series. While machine learning techniques for time series forecasting have been applied to various real-life problems, literature on its application to LVPS transactions is lacking. Thus, this study aims to apply machine learning techniques in forecasting the time series generated from transactions of individual LVPS participants.
Single hidden-layer feedforward networks (SLFN) trained using the extreme learning machine (ELM) algorithm were used as the machine learning model for forecasting the time series derived from the LVPS participants. In the model building process, Monte Carlo cross-validation (MCCV) was applied in making design decisions such as the number of hidden units, data preprocessing, outlier treatment, ensemble technique, and data disaggregation. The resulting ELM models were compared to an automatically generated ARIMA model. For models built using aggregated data, 15 out of 29 or 52% of the ELM models obtained the lowest error metric. For models built using disaggregated data, 6 out of 7 or 86% of the ELM models obtained the lowest error metric. These results show the potential of ELM neural network as an alternative model for time series forecasting applications. Multi-step forecasts were also generated comparing recursive, direct, and DirRec strategies with results pointing to the general applicability of the recursive strategy compared to other more complex multi-step forecasting strategies.
Further studies that can help bring ELM networks into eventual real-world deployment in the LVPS domain include studies on alternative means of setting the embedding dimension, consideration of other exogenous variables that may affect LVPS transactions, and use of other metrics or statistics aside from forecast accuracy as a guide for the model building process. |
---|