Context
The archive has no CBS statewide release for FY 2014/15, and the CBS website blocks automated downloads, so the twelve months were never transcribed. The LGA release for that year does exist and gives an annual NGR total for the state. In every year where both series exist, the LGA total matches the statewide total to within monthly rounding (the largest gap is $15,769 in FY 2016/17, about 0.002%).
Some of the 2026 analyses need a complete monthly series: STL decomposes an unbroken sequence, and an interrupted time series with autocorrelated errors is simplest without a hole in the middle.
Decision
The twelve months stay missing. The annual NGR chart shows FY 2014/15 as a hollow marker taken from the LGA release, labelled as such, and the statewide tables leave the year blank with a note. Year-on-year changes are not computed across the gap. The STL decomposition runs separately on July 2009 to June 2014 (60 months) and July 2015 to June 2025 (120 months). The interrupted time series uses July 2015 to June 2025 only, which also keeps its pre-closure period to the same recording regime.
Options considered
- Interpolate the twelve months. It would make the series look complete, but every chart and model would then rest on numbers nobody published.
- Spread the LGA annual total over months using an average seasonal pattern. Better than straight interpolation, but still invented monthly data, and the seasonal pattern would be fitted to the very series it fills.
- Use a method that tolerates gaps (for example a state-space model). Defensible, but heavier than the question needs, and hard to check against a reference implementation.
- Leave the gap and work around it (chosen).
Why
A visible gap is honest and costs little. The annual total from the LGA release answers the most common question about the year, and the two runs are long enough for the methods that need unbroken data.
What happened
The first run (60 months, five cycles) is short for STL: its seasonal estimate rests on five values per calendar month. Its trend and seasonal strengths (0.84 and 0.91) are higher than the later run’s (0.72 and 0.64), partly because the later run contains the COVID-19 years. The interrupted time series has 56 pre-closure months, enough to estimate a trend, but not enough to also fit a pre-existing slope change, so the model assumes the July 2015 to February 2020 trend was linear.
What I’d change
I would request the FY 2014/15 statewide release from CBS directly and add it through the same transcription and PDF check as the other years. If it arrives, this record will be superseded, and the STL and the interrupted time series will be refitted on one run from July 2009.