Testing a simple averaged model for local and

regional population forecasts

Tom Wilson

Introduction

Subnational population forecasting is challenging

Past forecast errors often high, especially age-specific forecasts

Less detailed, lower quality, and noisier base period data

High user expectations

Often hundreds or thousands of local areas and regions

Staffing and budget limitations

Introduction

Variety of local and regional population forecasting methods, e.g.

• extrapolation

• various forms of cohort-component model

• dwelling-led models

• regression models

• disaggregation approaches

• land use development and dwelling allocation models

• microsimulation

• large-scale population-housing-employment-transport etc. models

Not much attention given to combining, especially averaging

Aims

To assess performance of averaged model,

Constant Share of Population–Variable Share of Growth (CSP-VSG)

1. Does the averaged model generate short-term population forecasts for

a range of subnational geographies which are of acceptable accuracy and

more accurate than those from linear extrapolation?

2. For which areas were CSP-VSG averaged model forecasts successful

and where were they less successful?

Data & methods

Averaged model

Constant Share of Population (CSP) model

Variable Share of Growth model (VSG)

CSP: local population is fixed proportion of an independent State forecast

VSG: local population growth is varying share independent State growth

Retrospective forecasts created for:

average of models’ outputs

Geography 1991-2001 1996-2006 2001-2011

SA2 local areas

SA3 minor regions

SA4 major regions

Data & methods

Linear extrapolation

Comparative naïve forecasts

State and Territory population forecasts

Medium series population projections produced by the Australian Bureau

of Statistics used

Estimated Resident Populations (ERPs)

Forecasts compared with ERPs (official population estimates)

Error measures

Absolute Percentage Error

Median Absolute Percentage Error (MedAPE)

% of areas forecast with less than 10% APE

Results: average errors

0 2 4 6 8 10

State projections

State ERPs

Metro region ERPs

Non-metro ERPs

State projections

State ERPs

Metro region ERPs

Non-metro ERPs

State projections

State ERPs

Metro region ERPs

Non-metro ERPs

SA2

loca

lar

eas

SA3

min

or

regi

on

sSA

4 m

ajo

rre

gio

ns

MedAPE (%)

Averaged model

Linear extrapolation

Results: % areas with < 10% APE

0 20 40 60 80 100

State projections

State ERPs

Metro region ERPs

Non-metro ERPs

State projections

State ERPs

Metro region ERPs

Non-metro ERPs

State projections

State ERPs

Metro region ERPs

Non-metro ERPs

SA2

loca

lar

eas

SA3

min

or

regi

on

sSA

4 m

ajo

rre

gio

ns

% < 10 APE

Averaged model

Linear extrapolation

Results: average errors by population size

0 5 10 15 20 25 30 35

0-2,4992,500-4,9995,000-9,999

10,000-14,99915,000-24,999

25,000+

0-24,99925,000-34,99935,000-44,99945,000-59,99960,000-99,999

100,000+

0-149,999150,000-249,999

250,000+

SA2

loca

lar

eas

SA3

min

or

regi

on

s

SA4

maj

or

regi

on

s

MedAPE (%)

Averaged model

Linear extrapolation

Results: averaged errors by base period growth

Annual average

growth rate (%)

of 10 year

base period

0 5 10 15 20 25 30

< -1%-1 to 0%0 to 1%1 to 2%2 to 3%3 to 4%4 to 5%

5% +

< -1%-1 to 0%0 to 1%1 to 2%2 to 3%3 to 4%4 to 5%

5% +

< -0%0 to 2%

2% +

SA2

loca

l are

asSA

3 m

ino

r re

gio

ns

SA4

maj

or

regi

on

s

MedAPE (%)

Averaged model

Linear extrapolation

Results: average errors by base period growth volatility

Absolute difference

of annual average

growth rate of 1st

5 years of base

period &

annual average

growth rate of 2nd

5 years of base

period

0 5 10 15 20

0-0.10.1-0.5

0.5-11-22-33+

0-0.10.1-0.5

0.5-11-22-33+

0-0.50.5-2

2+

SA2

loca

lar

eas

SA3

min

or

regi

on

s

SA4

maj

or

regi

on

s

MedAPE (%)

Averaged model

Linear extrapolation

Results: average errors by metro/non-metro

0 2 4 6 8 10 12

Metro

Non-metro

Metro

Non-metro

Metro

Non-metro

SA2

loca

lar

eas

SA3

min

or

regi

on

sSA

4 m

ajo

rre

gio

ns

MedAPE (%)

Averaged model

Linear extrapolation

Results: comparison with Queensland projections

MedAPEs after 10 years by population size category

Jump-off

population

CSP-VSG

averaged model

Linear

extrapolation

Queensland

projections

100-2,499 20.1 30.2 10.7

2,500-4,999 8.2 11.5 7.3

5,000-9,999 7.5 9.3 7.4

10,000-14,999 5.9 7.9 6.5

15,000-24,999 5.1 6.6 6.4

25,000+ 4.3 5.4 6.2

All sizes 6.3 8.1 7.6

Results: modelling errors

State projection-constrained forecasts for SA2 local areas

Effect Forecast period

1996-2006 2001-2011

Intercept 73.218*** 61.658***

ln(population) -7.376*** -6.299***

Base period growth rate 0.378*** 0.312***

Base period growth rate volatility 0.269*** 0.960***

Metropolitan / non-metropolitan 3.873*** 3.992***

Adjusted R2 0.220 0.235

Key points

Relative to linear extrapolation averaged model produces

• lower average forecast errors, and

• greater proportion of areas forecast within 10% error

% point reduction in error greater for SA2 local areas

Non-metropolitan-constrained forecasts more accurate

Better than Queensland projections for 10,000+ people

Poor forecasts for some areas

Forecasts simple and easy to produce, requiring minimal data inputs, staff

time and organisational resources

Strengths & weaknesses of the averaged model

Strengths Weaknesses

Simple model: easy to understand Atheoretical

Low input data requirements Outputs total population only

Largely automated: ready-made Excel spreadsheet

template available

Cannot be applied to areas with zero population

at the start of the base period

Forecasts can be produced very quickly and

cheaply for hundreds or thousands of areas

Performs poorly for some areas undergoing large-

scale residential (re)development

Relatively low average errors demonstrated for

Australia over 10 year forecast horizons

Some areas will have large errors

Links to an independent forecast for a State/large

region or other ‘parent region’

Difficult to incorporate local area-specific

assumptions and alternative scenarios

Reduces decline of declining populations and slows

growth of rapidly growing populations

Possible uses of the averaged model

Averaged model could form a useful part of a subnational population

forecasting system.

1. Integral part of a subnational forecasting system for all, or just the non-

metropolitan part, of a State.

2. Validate forecasts from another model by providing an independent set

of forecasts.

3. Benchmark set of forecasts when undertaking retrospective tests of

other potential forecasting models.