|
CINTRAFOR |
|
Working Paper 76 |
|
Time Series
Methods for Commodity Price Forecasting: |
|
Gerard Alexander Malcolm |
|
94 pages. |
Introduction
PULP
(pulp) n.
1. A soft,
moist, shapeless mass of matter. 2. A magazine or book
containing lurid subject matter and being characteristically printed on rough,
unfinished paper.
American Heritage Dictionary, New College Edition (1993)
Many commodity products
have similar characteristics. They are relatively homogenous products that can
be produced using readily available technology, and are traded internationally on competitive markets. They are based on
natural resources, the availability of which is subject to shocks. They are
intermediate inputs with few short-term potential substitutes. As a result of
these characteristics, commodity prices tend to be volatile.
Market pulp is in many ways
a typical commodity. It also has some particular characteristics that
contribute further to price volatility. These include high capital intensity,
long-lived capital equipment, and speculative inventory management behavior on
the part of consumers. Because of this volatility, accurate forecasting of
market conditions is a difficult but potentially very useful exercise.
Investment decisions are made on the basis of price forecasts, and improving
their accuracy can lead to better decision making on the part of producers.
This may in turn reduce the volatility of prices.
Considerable energies are
devoted to the task of forecasting pulp prices. A number of recognized
short-term leading indicators for price movements exist, and these are widely
studied by industry participants. However, less attention is paid to forecasts
of the more distant future. This is surprising, given that the relevant time
horizon for some of the most important decisions made by the industry (namely,
investments in new capacity) is at least five years. This may perhaps be
explained by a relative dearth of useful long-term leading indicators on which
to base forecasts. The purpose of this research is to contribute to our
understanding of how the market functions and to our ability to forecast market
conditions, in both the short term and the long term.
-The first step in
achieving this is to identify the most important causal factors in determining
market conditions. This is done by means of a survey of existing literature and
discussion with industry experts. The focus of this survey is on prices and the
variables that interact with prices.
-The second step is to
assess how best to model these factors in such a way as to be useful for
forecasting purposes. Initially, it was intended that this be done using
structural methods, focusing particularly on forecasting investment levels (and
then prices). In the course of developing the research proposal, however, it
became apparent that recently developed time series methods provided a superior
means to this end.
Since the late 1980s, time
series econometrics has undergone something of a revolution. Two key concepts
are non-stationarity and cointegration. Broadly speaking, the first of these
refers to a variable which does not have a fixed trend, and the second refers
to two or more such variables moving ‘in parallel’ in the long term.
Non-stationarity occurs when a variable is subject to stochastic shocks that
have permanent effects on it. It has been recognized that many economic
variables (including commodity prices) do exhibit non-stationarity, and it has
been found that previously used methods based on the assumption of stationarity
are not valid for modeling such variables. Methods have been developed to test
for non-stationarity, and to model systems of non-stationary variables.
Cointegration methods
provide a means for the modeler to take advantage of the effective treatment of
short-term dynamics which time series models provide, while ensuring that
long-term forecasts have sensible properties. The concept of cointegration is
an intuitively sensible one. A pair of variables is said to be cointegrated if
they have a tendency to maintain a fixed ‘equilibrium’ relation to one another
over the long term. In a stable cointegrating relationship, the variables will
adjust to eliminate any divergence from this equilibrium. Such relationships
make a great deal of sense for several variables of relevance to the present
case, and forecasts for these variables which show significant and long lasting
violations of these relationships are not plausible. Cointegration methods
allow us to test for and model the existence of long-term equilibrium
relationships between the variables in a system, within the framework of a
dynamic time series model. A focus of this research is to assess whether these
methods prove to be of practical use for building forecasting models.
One difficulty encountered
in this research is that little previous time series research on pulp markets
has been published. We are therefore obliged to adopt an incremental approach
to model building. First, the time series characteristics of pulp prices
themselves are explored. We then develop a series of multivariate time series
models. These models will attempt to capture some of the short-term and
long-term processes that are thought to determine prices. The success or
failure of each model will be judged according to three criteria:
- Its statistical
acceptability as a representation of the data generation process;
- The plausibility of its
estimated coefficients;
- And its
out of sample forecasting performance.
Forecasting performance is
assessed both on the basis of a formal measure of forecast error, and on an
informal assessment of the desirability of the properties of the forecasts.
While these models are
designed for forecasting purposes, they may also be used to an extent for
structural analysis. This research is not designed primarily to test any
specific behavioral hypotheses, but where convenient, structural analysis will
be used to provide insights into how the market functions.
The structure of this
thesis is as follows. A description of the pulp market is provided in chapter
1. This outlines the short-term and long-term dynamics of the market. It also addresses
the issue of how the ‘market’ should be defined in terms of its product and
geographical scope. Chapter 2 provides a brief overview of the econometric
concepts mentioned above, along with a description of relevant modeling
methods. Data availability for the variables found to be of importance is
addressed in chapter 3. The core of the research, which consists of the
estimation and testing of several models of pulp prices and other variables, is
described in chapter 4. The forecasting performances of the different models
are compared in chapter 5. Chapter 6 concludes.