Analysis · Trends

How revenue moved, read with its uncertainty

The Statewide page shows what CBS published. This page asks three questions of the same monthly series: what is trend and what is season, how the series changed around the 2020 venue closures, and how much the yearly figures depend on month-to-month variation. Dollars are in average FY 2024/25 dollars (ABS CPI, Adelaide) unless stated.

Gambling Help Line 1800 858 858 or gamblinghelponline.org.au.

Section 1

Trend and seasonality (STL)

STL splits each month into a slowly changing trend, a repeating seasonal pattern and a remainder. The decomposition needs an unbroken series, so it runs separately on July 2009 to June 2014 (60 months) and July 2015 to June 2025 (120 months), either side of the missing FY 2014/15. The robust version gives low weight to months that don’t fit, so the 2020 closures don’t drag the trend down.

Monthly NGR, decomposed

Real NGR ($ million a month). Shaded: no statewide release (FY 2014/15) and the 2020 closures.

Monthly NGR and its trend

Monthly NGR
Trend
Set aside by the robust fit

Seasonal component

Remainder

R’s stl(x, s.window = 13, robust = TRUE), ported to TypeScript and checked against R to eight significant figures: seasonal window 13, trend window 21, low-pass window 13, 15 robustness iterations. Hollow markers: the months the robust fit gave a weight below 0.5.

Strength. On the later run, leaving out the four closure months, the trend explains 72% and the seasonal pattern 64% of the variation they share with the remainder (Wang, Smith and Hyndman’s strength measures, n = 116 months). On 2009–2014 the figures are 84% and 91%. These are descriptive: STL has no sampling model, so no interval is attached.

Months set aside. The robust fit gave a weight below 0.5 to 15 months of the later run, without being told about COVID-19: Dec 2017, Mar 2020, Apr 2020, May 2020, Jun 2020, Jul 2020, Aug 2020, Nov 2020, Dec 2020, Jan 2021, Feb 2021, Apr 2021, Jul 2021, Jan 2022, Dec 2023. They include all four closure months and both short lockdowns (November 2020 and July 2021).

Seasonal effect by calendar month, July 2015 to June 2025

Month	Mean effect	Range across years
Aug	+$6.7m	+$6.6m to +$6.8m
Jul	+$6.0m	+$5.6m to +$6.4m
Oct	+$2.3m	+$2.2m to +$2.5m
Sep	+$1.7m	+$1.6m to +$1.8m
May	+$0.2m	−$0.2m to +$0.5m
Mar	−$0.2m	−$0.4m to $0.0m
Dec	−$0.6m	−$0.8m to −$0.3m
Nov	−$0.7m	−$0.9m to −$0.6m
Jun	−$1.9m	−$2.2m to −$1.8m
Apr	−$2.7m	−$2.9m to −$2.6m
Jan	−$3.0m	−$3.1m to −$2.8m
Feb	−$7.7m	−$7.9m to −$7.4m

Part of the pattern is calendar length: February has two or three fewer days than the months around it. The range shows how the seasonal component drifts across the ten years (the seasonal window of 13 lets it change slowly).

Section 2

Before and after the 2020 closures

An interrupted time series fits one line to the months before the closures and lets both the level and the slope change at reopening: monthly NGR = level + trend · month + change in level after July 2020 + change in trend after July 2020 + a calendar-month effect. It uses July 2015 to June 2025 and leaves out March to June 2020, when venues were shut. Monthly errors are autocorrelated (lag-1 autocorrelation of the residuals 0.39), so the intervals use Newey–West standard errors, which allow for that.

Interrupted time series of monthly NGR

Choose a specification to see how the estimates move when the outcome or the months change.

Specification

Monthly NGR in FY 2024/25 dollars, July 2015 to June 2025, closure months left out.

Observed
Fitted model
Pre-closure trend carried forward

Change in level at reopening (July 2020): +$9.3m (95% CI +$5.8m to +$12.8m)
Change in trend, per year: +$3.4m (95% CI +$2.2m to +$4.6m)
Gap to the carried-forward trend, June 2025: +$26.2m (95% CI +$20.4m to +$32.0m)

116 months; Newey–West standard errors with lag 4; t intervals on n − 15 degrees of freedom. Values are per month.

Segmented regression with July as the reference month; Newey–West (Bartlett) standard errors with the rule-of-thumb lag floor(4 (n/100)^(2/9)); 95% t intervals. Checked against statsmodels in the unit tests.

All four specifications

Specification	Months	Level change at reopening	Trend change per year	Gap, June 2025	Residual lag-1 ACF
Real NGR (primary)	116	+$9.3m95% CI +$5.8m to +$12.8m	+$3.4m95% CI +$2.2m to +$4.6m	+$26.2m95% CI +$20.4m to +$32.0m	0.39
Nominal NGR	116	+$5.9m95% CI +$3.0m to +$8.8m	+$5.5m95% CI +$4.4m to +$6.5m	+$32.7m95% CI +$27.8m to +$37.7m	0.40
Real NGR, short lockdowns left out	114	+$11.0m95% CI +$7.9m to +$14.1m	+$2.9m95% CI +$1.9m to +$4.0m	+$25.5m95% CI +$19.9m to +$31.1m	0.60
Real NGR per machine	116	+$1,05795% CI +$797 to +$1,318	+$21395% CI +$113 to +$313	+$2,10495% CI +$1,631 to +$2,577	0.36

Per-machine rows are dollars per machine per month. The gap is the fitted value minus the pre-closure trend carried forward to June 2025 (counterfactual $56.7m a month for the primary specification).

What the estimates do and don’t say

In every specification NGR came back above its pre-closure path and kept rising faster than before. The model describes that break; it can’t say why. Income support, other entertainment closing, interstate borders and inflation all changed in the same months, so none of these figures is the effect of the closures alone.

Why real dollars are the primary outcome

Prices rose quickly from 2021. In nominal dollars part of that inflation shows up as a steeper post-2020 trend (+$5.5m (95% CI +$4.4m to +$6.5m) a year, against +$3.4m (95% CI +$2.2m to +$4.6m) in real terms). See DR-005 for the choices behind the model.

Section 3

Like-for-like years, month by month

A simpler check that needs no model: pair each calendar month of one financial year with the same month of another, so seasonality cancels out, and look at the mean difference. Intervals come from the t distribution and from 4,000 bootstrap resamples of the twelve pairs (seed 20090701).

Comparison	Before (mean)	After (mean)	Mean paired difference (t CI)	Bootstrap CI	d_z	p
Real NGR: first full year after reopening against the last full year before COVID-19FY 2018/19 → FY 2020/21, n = 12 months, +9.3%	$70.95m	$77.56m	+$6.61m (95% CI +$2.44m to +$10.78m)	+$2.68m to +$9.87m	1.01	0.005
Real NGR: FY 2024/25 against FY 2018/19FY 2018/19 → FY 2024/25, n = 12 months, +18.5%	$70.95m	$84.05m	+$13.10m (95% CI +$12.05m to +$14.16m)	+$12.24m to +$14.04m	7.89	< 0.001
Real NGR per machine: FY 2024/25 against FY 2018/19FY 2018/19 → FY 2024/25, n = 12 months, +22.4%	$5,848	$7,156	+$1,308 (95% CI +$1,219 to +$1,397)	+$1,236 to +$1,388	9.33	< 0.001

d_z is the mean paired difference divided by the standard deviation of the differences. With only twelve pairs a percentile bootstrap interval tends to be a little too narrow, which is why the t interval is shown first. Both intervals and the p-value treat the twelve monthly differences as independent; neighbouring months are correlated (the same trend and the same shocks), so the intervals are probably somewhat too narrow. Read them as a description of the twelve pairs, not as a precise test.

Section 4

NGR per machine, with intervals

Annual NGR per machine is a ratio of twelve months of revenue to the year’s average machine count. Resampling the twelve months (with replacement, 2,000 times, seed 20090701) shows how much the figure leans on particular months. The point estimates are exactly the Statewide page’s.

Annual NGR per machine

Dots: the published-basis figure. Bars: 95% percentile bootstrap intervals over months.

Dollars

FY 2014/15 has no statewide release. FY 2019/20 is left blank, as on the Statewide page: March to June 2020 carry NGR against zero reported machines.

Annual NGR per machine by financial year with 95% bootstrap intervals over months, in FY 2024/25 dollars and in nominal dollars
Financial year	Months	Real, FY 2024/25 $ (95% CI)	Nominal $ (95% CI)
FY 2009/10	12	$85,698 ($82,687–$88,867)	$57,305 ($55,403–$59,282)
FY 2010/11	12	$84,751 ($81,483–$88,124)	$58,463 ($56,472–$60,455)
FY 2011/12	12	$82,461 ($79,880–$85,150)	$58,425 ($56,594–$60,327)
FY 2012/13	12	$80,131 ($77,518–$82,966)	$57,919 ($56,081–$59,894)
FY 2013/14	12	$78,447 ($75,379–$81,475)	$58,144 ($56,049–$60,216)
FY 2014/15	No statewide release archived
FY 2015/16	12	$76,605 ($74,482–$79,096)	$58,181 ($56,566–$60,059)
FY 2016/17	12	$71,896 ($69,099–$74,778)	$55,424 ($53,350–$57,555)
FY 2017/18	12	$70,983 ($68,896–$73,047)	$55,962 ($54,440–$57,429)
FY 2018/19	12	$70,166 ($67,764–$72,554)	$56,179 ($54,360–$57,972)
FY 2019/20	Not computed: NGR recorded against zero machines (COVID-19 closures)
FY 2020/21	12	$81,201 ($76,100–$86,022)	$67,171 ($62,986–$70,985)
FY 2021/22	12	$82,491 ($78,997–$85,385)	$71,150 ($67,879–$73,902)
FY 2022/23	12	$84,495 ($81,217–$87,882)	$78,591 ($76,097–$81,115)
FY 2023/24	12	$83,659 ($81,896–$85,273)	$81,664 ($80,023–$83,114)
FY 2024/25	12	$85,869 ($83,520–$88,091)	$85,858 ($83,649–$87,976)

2,000 percentile-bootstrap resamples of the months in each year, seed 20090701. The point estimates are the Statewide page’s figures.

Intervals on a census

CBS reports every machine and every dollar, so there is no sampling error in the usual sense. The intervals on this page describe how stable a figure is given the month-to-month variation in the series: a wide interval means a few unusual months move the annual figure, not that the figure was estimated from a sample.