Business and Economic Forecasting Unit – Current Projects
Exponential smoothing using a state-space framework
Researchers in the unit have been responsible fore developing a statistical
framework for the exponential smoothing methods of forecasting over the last
few years. The framework facilitates the generation of logically sound prediction
distributions. It also permits the use of information criteria for choosing
between competing methods. Work is continuing on refinements to the approach.
KEY RESEARCHERS: Ralph Snyder, Rob Hyndman, Muhammad Akram, Anne Koehler,
Keith Ord, Ashton de Silva.
Time series of counts
Forecasting methods for time series data consisting of counts are being developed
in this project. Such time series arise in many contexts including intermittent
demand and stock transaction data. The aim is to extend the range of available
models to allow covariates to be included linearly or non-parametrically and
to provide a range of serial correlation patterns.
KEY RESEARCHERS: Gael Martin, Ralph Snyder, Rob Hyndman, Lydia Shenstone,
Chris Strickland, Lucy Gunn.
Functional demographic forecasting
Some data comes in the form of functions rather than individual observations.
For example, mortality is a function of age, and it is of interest to forecast
mortality in order to predict future population profiles. Another context
for this type of data is in cancer monitoring where a person’s cancer status
is predicted based on functional biometric data. In this project, we aim to
develop new methods for predicting functional data, especially as applied
to age-specific demographic forecasts.
KEY RESEARCHERS: Rob Hyndman, Shahid Ullah, Don Poskitt, Arivulzaham Sengarapillai,
Heather Booth (ANU), Leonie Tickle (Macquarie)
Hierarchical forecasting
In this project we are forecasting large numbers of related time series which
can be aggregated at several different levels. Important applications are
to the Pharmaceutical Benefits Scheme and Labour Market forecasting. We aim
to develop new statistical methodology for forecasting hierarchical time series
which (1) provides point forecasts that are consistent across the levels of
hierarchy; (2) allows for the correlations and interaction between the series
at each level of the hierarchy; (3) provides estimates of forecast uncertainty
which are consistent across the levels of hierarchy; and (4) is sufficiently
flexible that ad hoc adjustments can be incorporated and important covariates
can be included.
KEY RESEARCHERS: Rob Hyndman, Chandra Shah, Roman Ahmed.
Model selection for automatic forecasting
When forecasting large numbers of items, it is important to have a fully
automated procedure. This project is looking at automatic model selection
for forecasting, in the context of exponential smoothing models and ARIMA
models.
KEY RESEARCHERS: Rob Hyndman, Ralph Snyder, Yeasmin Khandakar
Environmental and epidemiological forecasting
Some respiratory diseases are impacted by environmental factors such as pollutant
levels, pollen levels and meteorological conditions. In this project, we are
developing some nonparametric models of the relationship between asthma hospital
admissions and the various environmental covariates. This can then be used
to predict hospital admissions a few days in advance which enables hospital
staffing levels to be adjusted accordingly. A related task is to develop models
of the covariates themselves so they can be forecast separately.
KEY RESEARCHERS: Rob Hyndman and Bircan Erbas (University of Melbourne).
Forecasting in the frequency domain
This project looks at a frequency domain forecasting method and compares
it to conventional time domain forecasting methods. The use of this virtually
unknown frequency domain method is highlighted in amongst the vast, almost
exclusive time domain forecasting literature of business, financial and economic
applications.
KEY RESEARCHERS: Ann Maharaj
Wavelet multiscale forecasting
In this project, a procedure to obtain forecasts for stationary time series
and stationary long memory time series by means of wavelet multi-resolution
analysis has been developed. Outcomes so far for stationary long memory processes
are very encouraging. Work is continuing to further refine the procedure.
KEY RESEARCHERS: Ann Maharaj and Don Percival (University of Washington)
Tourism forecasting
Funded by Tourism Research Australia and the CRC for sustainable tourism,
this project is developing new methodology for forecasting tourist numbers.
In particular, we are developing methods that can forecast domestic tourist
numbers within Australia as well as tourists arriving and leaving Australia.
KEY RESEARCHERS: George Athanasopoulos and Rob Hyndman
Multivariate modelling and forecasting
There are several advantages in analysing and modelling time series jointly.
Multivariate models are able to explore and use information that emanates
from dynamic inter-relationships amongst variables. For example, a group of
series may be contemporaneously correlated, or one series may be a leading
indicator to several others, or feedback relationships may exist. The flexibility
and capacity of multivariate models to utilise this type of information can
lead to improved analysis and improved forecasting performance. In this project
we are aiming to develop methodology for estimating multivariate ARIMA and
exponential smoothing models.
KEY RESEARCHERS: George Athanasopoulos, Ashton de Silva, Farshid Vahid (ANU),
Ralph Snyder, Rob Hyndman, Don Poskitt.
|
