At one point, my Australian population projection said the country had roughly 52 million people. The expected total was closer to 26 million. Unless I had accidentally projected several decades into the future, something had gone badly wrong.

I was projecting population at a very granular geographic level. The original method could produce reasonable-looking results for individual areas, but small-area growth is noisy. A new development, a low starting population, or an unusual recent change can create a growth rate that should not be extended too confidently.

Applied independently across thousands of areas, those local exaggerations accumulated. Most rows still looked plausible on their own. It was only when I added everything back together that Australia had somehow doubled in size.

Australia already has enough infrastructure issues serving about 26 million people. My model inventing another 26 million residents was probably not going to improve matters.

I replaced the original approach with a simpler linear projection model. It gave up some apparent sophistication at the local level, but the resulting totals were more credible, easier to explain, and much easier to validate against broader population benchmarks.

The lesson was not that linear projection is always the right answer. It was that granular models can fail in ways that still look convincing row by row. Whenever I project small-area data now, I roll it back up to a known total. A model that cannot survive that check does not deserve to make it into a dashboard.