Statistics and the future

Statistics are a reliable guide for planning where the underlying system is stable and the variable is well-measured, but they become unreliable precisely when planning matters most, in the face of structural breaks and rare high-impact events, because the data records a past that the future has stopped resembling.

• Statistics are reliable for planning when the system generating them is stable and the question is narrow.
  Why Where behaviour repeats, the past is a good sample of the future, so demographic and actuarial trends let a state build the right number of schools, beds and trains years ahead.
  Example Singapore's population planning runs on demographic statistics: a total fertility rate of 0.97 in 2024 and a citizen population projected to shrink from the early 2040s drive housing, immigration and eldercare policy decades out (SingStat, as_of 2026-06).
  Then evaluate But even here the number guides rather than dictates: the same fertility data has not told the state which pro-natalist lever actually works.
• Statistics turn unreliable when the system itself breaks, because the data sample no longer represents the world being planned for.
  Why A structural break means past relationships dissolve, so a forecast extrapolated from pre-break data plans for a world that has ended.
  Example No pre-2020 epidemiological model based on ordinary flu seasons prepared most states for COVID-19's spread, and pandemic projections swung wildly as the system kept shifting (global pattern, as_of 2026-06).
  Then evaluate The honest limit: the rarer and more consequential the event, the thinner the data, so statistics fail exactly where the stakes are highest.
• Statistics also mislead when the measure is mistaken for the thing, so planning optimises the proxy instead of the goal.
  Why Once a number becomes a target, people game it, so the statistic stops measuring what it once did and the plan built on it drifts from reality.
  Example GDP growth long stood in for national wellbeing until rising output masked stagnant median wages and ecological cost, prompting the search for broader indicators (global pattern, as_of 2026-06).
  Then evaluate This is Goodhart's trap: a reliable measure becomes unreliable the moment a plan is built to maximise it.
• The deeper unreliability is that statistics describe the past and planning shapes a future people can change.
  Why Human responses to a forecast can falsify it, so a statistic that warns of disaster may avert the very outcome it predicted, or a rosy one may breed the complacency that undoes it.
  Example The Club of Rome's 1972 resource-collapse projections did not arrive on schedule, partly because the alarm helped spur the conservation and substitution the models had not assumed (widely documented, as_of 2026-06).
  Then evaluate The paradox the question must face: the best statistical guide to the future is one acted on so well that it stops coming true.

The question is mis-posed: statistics are neither reliable nor unreliable in themselves, because they are not the guide, they are the raw material the guide is made from, so the real variable is the quality of human interpretation, and asking whether statistics are reliable is like asking whether bricks are a reliable house.

• The same statistic supports opposite plans depending on the question asked of it, so reliability lives in the framing, not the figure.
  Why A number is mute until a person decides what counts, what to compare it to and what to ignore, so two analysts can read one dataset toward two conclusions without either lying.
  Example Singapore's low recorded crime can be read as proof of effective policing or as a product of tight social control, the same statistic carrying two stories (SPF data, as_of 2026-06).
  Then evaluate So the failure attributed to 'unreliable statistics' is usually a failure of the question put to them.
• Most famous statistical failures are failures of selection and assumption, not of the numbers.
  Why When the sample is biased or the model assumes the wrong relationship, the output is wrong even if every figure is accurate, so the error is human before it is numerical.
  Example The 2008 financial crisis was forecast away by models that assumed housing-default risks were uncorrelated, an assumption error dressed as statistical confidence (widely documented, as_of 2026-06).
  Then evaluate The complication for the question: the data was not lying, the modellers' premise was, which relocates the problem entirely.
• Reframed this way, the right question is how to build the judgement that makes statistics usable, not whether to trust them.
  Why Good planning pairs data with scenario thinking, sensitivity to uncertainty and a sense of what the numbers omit, so the statistic becomes one disciplined voice rather than an oracle.
  Example Singapore's coastal-protection planning treats sea-level projections as a range to hedge against, committing to long-run adaptation spending rather than betting on a single forecast (PUB / national adaptation plans, as_of 2026-06).
  Then evaluate This is the usable lesson: treat statistics as a guide to the range of possible futures, not a prediction of the one that will arrive.
• Even perfect statistics cannot decide the value questions that planning ultimately turns on.
  Why Numbers can tell you the cost of an ageing population but not how to weigh the young against the old, so the most important planning choices are normative, where statistics have no vote.
  Example Singapore's choices about how to fund eldercare against childcare are guided by demographic data but settled by values about fairness across generations (national budget debates, as_of 2026-06).
  Then evaluate The reframe's payoff: blaming statistics for bad plans excuses the human judgement that is actually in charge.