We Graded Our Own Forecasts

By Wes Calder. From MFGCalcs Magazine, Issue 1, November 2026, page 59.

Our forecast for Brent crude in September, made from data through August, was $77.25 a barrel. Brent averaged $114.16, a miss of about a third, above the top of the forecast’s range. We had diesel at $4.11 a gallon; it averaged $6.29. We expected both prices to fall that month. Both rose, and diesel finished outside its range too.

Natural gas missed by about as much the other way and still counted as a hit: its range was wider than the forecast itself. The 10-year Treasury yield missed by about half as much and counted as a miss, outside a range about a percentage point wide; none of the five 10-year forecasts graded so far has landed inside its range, and all five came in low. A range wide enough to hold a miss of a third makes a hit cheap.

Each month, the forecasts we publish go into a ledger with their ranges, and each is graded once its month is in the data. Since the ledger opened in July, it has logged 2,712 forecasts, and 664 have come due. The ranges are drawn to catch eight outcomes in 10. So far they have caught seven, and two in three of the graded forecasts came in low. The steadier series held: in September our forecasts for four producer price indexes of metals and machinery, and for the emergency oil reserve, came within about 1% and inside their ranges.

January’s eight-day spike in spot natural gas at Henry Hub reached our model as a single figure, the average of the month’s daily prices. It learns from each series’ own monthly history and nothing else: no weather, no storage figures, no news. In tests on six months it had not seen, April through September, our Henry Hub forecast missed by an average of about 31%, about four times the average over every such test of the 114 series we forecast. None of the six was January or February.

The model, a form of exponential smoothing, carries a series’ level and trend forward, so it falls behind a series that keeps climbing faster than it has before. That describes the method, not why any one forecast missed. In tests on past months the model called the direction of the next month’s move right 53.2% of the time, barely better than a coin toss.

Our forecasts reach six months ahead, but none has yet been graded beyond three, and on the chart of misses by distance a three-month forecast typically missed by nearly twice as much as a one-month one.

The ledger also keeps a mistake of ours. All 12 daily and weekly series in the July forecasts had been anchored on July itself, a month still in progress. We changed that: from August on, those series have been anchored on the last complete month. The July forecasts stayed in. The ledger only grows, so no forecast in it can be revised after the fact.

The mistake is no excuse for September. The September forecasts for Brent, diesel, gas and the 10-year yield were anchored on August, a complete month by the time we made them.

So read our ranges as the width of past misses, not as odds. For most series the latest six months are held back to test the model, and each range is drawn from its misses on the months before, so this summer’s misses are not yet in those ranges’ widths. A price pushed by something the model cannot see can go straight through one.

Our October forecasts are already in the ledger. Next Quarter Watch, on Page 6, sets four of them against October’s readings through Oct. 14 to Oct. 20, depending on the series, and three are averaging above their ranges: Brent crude, diesel and the emergency oil reserve. The one inside is natural gas, and its range is once again wider than the forecast itself.

Live data from this story

• Where Our Ranges Held: Forecast accuracy

Read it in the magazine

• MFGCalcs Magazine, Issue 1, November 2026, page 59