every year the social security trustees publish the annual report, which contains 75-year financial projections for the program. this is a document that projects the US economy to the year 2100. i find this practice genuinely interesting — not because the projections are useful in any normal sense, but because we've built a major policy debate around numbers that are, without being unkind about it, essentially made up.
the report is rigorous. it makes explicit assumptions about fertility rates, immigration levels, GDP growth, labor force participation, interest rates, and life expectancy. the actuaries who produce it are doing their jobs carefully. the problem is not the quality of the work. the problem is that the task is impossible, and carefully produced numbers that cannot be right create a false precision that is worse than honest uncertainty.
the CBO regularly misses GDP growth by more than a full percentage point one year out. in 2019, essentially no forecasting model predicted the 2020 pandemic recession. in 2006, almost no mainstream forecast predicted the 2008 financial crisis. these aren't cherry-picked failures; they represent the general state of economic forecasting at short time horizons. forecasting one year out is extremely hard. forecasting seventy-five years out is, not to put too fine a point on it, impossible. the system has too many variables, too many feedback loops, too many unknown unknowns, and too many potential discontinuities. any model that projects 75 years forward is not a forecast in any meaningful sense; it's a scenario.
consider what 75 years actually means. the 75-year forecast published today covers the period from now to 2100. the 75-year forecast published in 1951 covered 1951 to 2026. try to imagine what a 1951 economic forecast of today would have said. it would have missed the computer revolution entirely. it would have missed the internet, the service economy transition, the rise of China as an economic power, containerized shipping and its effects on global trade, the collapse of soviet communism, the demographic effects of the baby boom, the expansion of women in the workforce, the rise and fall and rise of manufacturing automation. these are not edge cases. these are the central facts of the economic history of the period, and none of them were forecastable 75 years in advance. nobody had a model that would generate them. the inputs weren't available. the outputs couldn't have been predicted from the inputs even if they had been.
this doesn't mean the social security actuaries should stop publishing long-range projections. the alternative — "we have no idea what will happen in the long run, so we won't model it" — is worse for policy purposes. the projections are useful not as predictions but as scenarios: if current trends continue and no major unexpected events occur and the assumptions we're making now are roughly right, here is roughly what the financial picture looks like. this is a useful input to the policy conversation. the trust fund exhaustion date (currently projected for 2033 or thereabouts) is a useful political focal point even if the specific year is not to be taken literally.
the problem is that people don't use the projections as scenarios. they use them as predictions. senators hold hearings about whether the trust fund exhausts in 2033 or 2035 as if that two-year difference had meaningful precision. policy debates are structured around specific projected shortfall percentages as if those percentages were real. the ritual of precision — the specific numbers, the tables projected out to 2100, the actuarial methodology — creates an impression of certainty that doesn't exist and that distorts the policy conversation.
the honest version of the 75-year report would say something like: "the trust fund will likely face depletion at some point in the next 15-30 years under plausible assumptions about demographics and growth, and probably requires some combination of revenue increases and benefit modifications to remain solvent on a long-term basis, but the specific numbers projected here are scenario estimates, not forecasts, and should be treated accordingly." this is less legible. it does not permit senators to have arguments about specific trust fund exhaustion dates. but it would be more accurate. precision sells. uncertainty doesn't. so we get the precision, and we pretend it's something other than storytelling with actuarial tables.
philip tetlock's work on forecasting accuracy is the best empirical treatment of this I've found — the core finding being that even the best human forecasters degrade rapidly past 5-10 year horizons, and that aggregate prediction markets outperform expert panels at short to medium term forecasting. at 75 years, we're so far past the range where any model has predictive validity that the exercise is categorically different from forecasting. read the social security trustees report. the assumptions appendix is fascinating precisely because you can see how many essentially unforecastable variables the model requires you to assume.