Understanding Data Revisions

Benchmark revisions redraw a whole year

Beyond the routine monthly corrections, most major series undergo an annual benchmark revision against administrative records. These are larger and they are retrospective: they can move a full year of employment growth by several hundred thousand jobs, or shift the level of GDP for several quarters at once.

The practical consequence is that any chart of the last two years you look at today is not the chart people were trading at the time. Backtesting a strategy on revised historical data and then running it live on first prints is comparing two different things, and it is one of the quieter reasons a macro-driven system performs worse in production than in testing.

The market trades the first print regardless

None of this stops the initial release from being the moment of maximum volatility. It is the first available information, it arrives at a known second, and the algorithms that react to it do not wait for confirmation. The second and third estimates typically produce a fraction of the reaction, because most of their content is already in other data by then.

So the first print is what you trade and the revised figure is what was true. Those are two different jobs. Treating the initial release as a trading event and the revised series as the economic picture keeps them from contaminating each other, which is what happens when someone builds a multi-week directional thesis on a single advance estimate.

Read the series, size for the error bar

Two habits follow from this. The first is to look at the three-month moving average of any noisy series rather than the latest point. Payrolls averaged over three months absorbs both the monthly noise and the revisions, and it is what most policy commentary is actually referring to when it describes the labour market as cooling or firming.

The second is to treat the published number as a point estimate with a real error term attached, not a measurement. The household survey underlying the unemployment rate carries a 90% confidence interval of roughly 600,000 on the monthly change in employment, which is wider than almost every monthly change it reports. Even the far larger payroll survey sits near 130,000. A move in a series smaller than its own error bar is not information.

If you are positioned into a release, that error term is an argument for size rather than for conviction. A number this uncertain deserves a position you can be wrong about twice without damage, which is a risk management decision and not an economic one. The drawdown calculator is a blunt way to see what a run of those errors does to an account, and the arithmetic is unkind: a 30% drawdown needs a 43% gain to get back to flat.