Market Data

We Graded Our Own Forecasts. Here's What's Actually Predictable in Manufacturing.

Most forecasts are published and never scored. We scored ours: 113 series, 676 forecast observations, graded against what actually happened. The honest finding is not how good the forecasts are. It is how sharply predictability varies by what you are trying to predict.

Almost nobody grades their own forecasts in public, which is exactly why forecasting has the reputation it does. So we did. Across 113 forecast series and 676 scored observations as of Jul 2026, the median absolute error is 5.2% and the direction-hit rate, how often the forecast simply got up-versus-down right, is 56.6% over 553 directional calls. Neither number is the point. The point is the spread underneath them: some manufacturing series are genuinely predictable, others are close to coin flips, and pretending otherwise is how forecasts lose people's trust.

What forecasts well, and why

The most accurate forecasts cluster in one place: slow-moving, heavily-sampled activity and labor series. The single most accurate category is Factory activity, with a median error of just 0.5% across 12 observations. At the series level, the standouts include Manufacturing Average Weekly Hours, PPI: Iron and Steel Castings, Manufacturing Average Hourly Earnings, All Employees, all landing within a fraction of a percent. The reason is structural: these series have strong momentum, large samples, and little day-to-day volatility, so tomorrow looks a lot like today and a forecast that respects the trend is usually right. Predictability is not a property of the forecaster; it is a property of the series.

What doesn't, and why that's the honest part

At the other end, Fuels and spot energy carries a median error of 20.1%, an order of magnitude worse than the best category. These are the volatile, thinly-sampled, shock-driven series, the ones that move on news rather than momentum. A forecast for them is worth less, and the honest thing is to say so and to publish a wide uncertainty band rather than a confident line. A forecasting record that only advertised the easy wins would be marketing. One that shows the hard cases getting hard is a measurement, and only the second kind is worth building decisions on.

Predictability is not a talent. It is a property of the series. Some numbers tell you where they are going; others only tell you where they have been.

How to use a graded record

The practical value of grading forecasts is that it tells you how much weight to put on each one. Lean hard on the forecast for a high-momentum, low-error series like capacity utilization or the wage; treat a forecast for a shock-driven, high-error series as a rough scenario, not a plan, and widen your buffers accordingly. The same discipline applies to any forecast a supplier or consultant hands you: ask for the track record, by series, including the misses. If they cannot produce one, you are being sold a line, not shown a measurement. For the full graded record as it updates, the forecast pages keep the live scorecard.

What a forecastable series looks like

The 54-year record shows capacity utilization making a full round trip, which is why point-in-time comparisons mislead so badly here. Its high came at the close of 1973 around 88.21%, gave way over the following years to 67.32% by the end of 2009, and has climbed since to 75.56%. That leaves it 14% below the peak and well off the floor, so whether today looks high or low depends entirely on which year you anchored to.

The pulse forecast pages publish the accuracy track record, including the misses, by series and category. See the graded record

Published 2026-08-06.