### Abstract:

lnterest rate risk is one of the most important types of risk to which banks are
inherently exposed. lnterest rates determine a bank's profitability and have an effect
on a bank's liquidity and investment portfolio. It is, therefore, extremely important to
be able to predict interest rates accurately and manage interest rate risk effectively.
In trying to manage interest rate risk, banks rely on Asset and Liability Committees
(ALCOs). They also make use of several strategies, which are described (Gap,
Earnings Sensitivity Analysis, Duration Gap and Market Value of Equity sensitivity
analysis). The first step for these strategies, on which later steps depend, is to make
interest rate forecasts.
Forecasting plays such a crucial role because many significant decisions depend on
the anticipated future values of specific variables. Forecasts may be produced in
various different ways. The method chosen depends on the reason for and the
importance of the forecasts as well as on the costs of alternative forecasting
methods.
In an attempt to manage interest rate risk by being able to predict the next rates
correctly, several different models are used to try and predict interest rates for two
data sets, namely: BA (Bankers' Acceptances, which is money market data) and Esc
(Eskom, which is capital market data). They each have their place in the South
African financial system, which is described in general.
The chosen simple forecasting models that are used are: naive, moving average and
exponential smoothing models. The aim is to try to predict the direction of the next
interest rate (UP, CONSTANT, or DOWN) while supplying a point prediction of the
next rate (one-step ahead). The "best" simple forecasting models are determined by
specific set criteria (percentage of correct direction predictions, mean squared error
and tracking signals).
For the same time series, more advanced models are taken into account where the
aim is to try to find an interval wherein the future interest rates (not only in the short-term
but in the longer-term as well) are most likely to lie, using models based on the
data, as well as first differences. For the long-term forecasts, two types of more
advanced models are used, namely: Box-Jenkins models (where, specifically,
nonseasonal second-order autoregressive or AR(2) models are examined); and
volatility models that are found using a new technique that creates an interval by
using different volatility estimates.
The word 'volatility' used throughout the study refers to models with a fixed volatility
function and not dynamic volatility as in models such as the ARCH and GARCH
types. In this study, the range from simple to more complex time series models with
constant volatility are considered. The former, simple models and AR(2) models are
referred to as forecasting models, the latter more advanced models are referred to as
volatility estimates.
Short- and long-term predictions are, thus, made for each time series, at different
specifically chosen points. A comparison of the effectiveness of the forecasting and
volatility models is made.