# Forecast Interbank Payment Flow in Large Value Payment System

AI and Robotics

Oct 19, 2013 (4 years and 8 months ago)

120 views

Forecast Interbank Payment Flow in

Large Value Payment System

Problem Provider: The People

s Bank of China

-
China Problem Solving Workshop in Finance

Weihai, China, October 2008

Problem Solving Report

Problem

Background

Large Value Payment System (LVPS)

——

a major payment infrastructure in China

Participants: most of the banking institutes

Amount: over 90% of payment

Liquidity risk of LVPS

Closing time of money market: 4:00PM

Closing time of LVPS: 5:00PM

Participants: delay settlement for the reserve
requirement

Risk: system traffic for too many transactions;

shortage of liquidity of a participant

PBC should monitor and remind the participants

Problems

How to forecast the interbank payment
flow in LVPS

Analyze the statistical property of each
commercial bank

s payment flow

Build models based on bank

s own payment
flow

Forecast results should have higher
accuracy

Solving Group

Pr. Zhongxing Ye Shanghai Jiao Tong University

Pr. Yong Zhou Chinese Academy of Science

Pr. Gang Wei Shandong University

Pr. Zhijun Zhao Chinese Academy of Social Sciences

Dr. Ying Bao Industrial and Commercial Bank of China

Min Zhang Chinese Academy of Science

Chen Su/Jing Zhao/Liping Zhu Shandong University

3 days hard working!

Solving report

Outline

Property of Payment Flow

Statistic Models

Wavelet Neural Network Model

ARIMA
-
GARCH Model

Additive Time Series Model

Further Research

Property of Payment Flow

1

Two stages

Reason: Several times of the reserve requirement

ratio adjustment since July,2006

Object: number of daily paid
-
out transactions

Property of Payment Flow

2

Yearly Effect

Reason: New year is the end of a settlement year;

Springing Festival is the real end of a year in China

Property of Payment Flow

3

Monthly Effect

Reason: Bank checks the performance of the

employee at the end of the month

Property of Payment Flow

4

Weekly Effect

Fact: Employee always works hard at Tuesday

Statistic Models

Wavelet Neural Network Model

ARIMA
-
GARCH Model

Additive Time Series Model

Wavelet Neural Network

Artificial Neural Network

is a new computing technology. It has been widely used
because of its flexible model and less data requirement. ANN
is composed of interconnected units or artificial neurons.

Wavelet Analysis

is a kind of method breaking down data at different frequency
and combines time
-
domain analysis and frequency
-
domain

Wavelet neutral network

combines ANN and wavelet analysis. It processes both self
-
learning ability of ANN and local property of wavelet
transformation and has higher approaching ability, faster
convergence speed and better forecast results.

Wavelet Neural Network

Fitting and Forecast:

ARIMA
-
GARCH Model

ARMA model

for stationary series

ARIMA model

for nonstationary series

covert nonstationary data to stationary by difference

GARCH model

for conditional eteroskedasticity
series

Time Series Model

Forecast :

Additive Time Series Model

Build model in every month. One day in a month:

Simple Form:

where

Additive Time Series Model

Additive Time Series Model

For month:

LSE:

Additive Time Series Model

Different deta for different week:

Additive Time Series Model

Using average deta:

Additive Time Series Model

Seasonal Effect:

Additive Time Series Model

Trending Effect:

Additive Time Series Model

Fitting:

Additive Time Series Model

Forecast:

Further Research

Forecast:

Number of paid
-
in transactions

Amount of paid
-
out transactions

Amount of paid
-
in transactions

Thanks !