Route 21 · 76 winters of US heating demand

Winter folklore, tested

Every October the same signs come out: Siberian snow, the QBO, the Blob, a cold October. We scored ten of them against the winter's gas-weighted heating degree days, 1951–2026, and again as real forecasts made only from earlier winters.

10 of 10October indicators no better than random numbers

0that beat the 10-year normal as a forecast

1signal that survives: El Niño strength (r = −0.28, p = 0.02)

The evidence

76 winters · gas-weighted HDD

Correlation of each indicator, as known by 1 November, with that winter's US gas-weighted heating degree days (Nov–Mar) above or below their long-term trend. Lines are 95% intervals. The grey band is where 95% of random, meaningless predictors land (±0.23). Only El Niño strength clears it.

Each winter from 1970/71 forecast using only earlier winters: trend plus the indicator, scored against the 10-year normal that storage desks plot. Negative means the indicator made the forecast worse. Random predictors beat the 10-year normal only 2% of the time.

The myths

one card each

Busted

October Arctic Oscillation

“A negative AO in October sets up a cold winter.”

Correlation, Nov–Mar: −0.05
Dec–Feb: −0.02
vs 10-yr normal: −6%

Busted

October AO + NAO

“If the Atlantic blocks in October, it blocks all winter.”

Correlation, Nov–Mar: −0.03
Dec–Feb: +0.00
vs 10-yr normal: −5%

Busted

Siberian snow in October

“Fast October snow over Siberia means a cold eastern US.”

Correlation, Nov–Mar: +0.14
Dec–Feb: +0.16
vs 10-yr normal: −4%

Busted

October snow advance rate

“It's how fast the snow spreads, not how much there is.”

Correlation, Nov–Mar: +0.08
Dec–Feb: +0.09
vs 10-yr normal: −8%

Busted

The Pacific "Blob"

“Warm water off the West Coast builds the ridge that sends cold east.”

Correlation, Nov–Mar: −0.02
Dec–Feb: −0.09
vs 10-yr normal: −7%

Busted

QBO at 30 hPa

“Easterly QBO, cold winter.”

Correlation, Nov–Mar: −0.11
Dec–Feb: −0.07
vs 10-yr normal: −7%

Busted

QBO at 50 hPa

“Same idea, lower in the stratosphere.”

Correlation, Nov–Mar: +0.02
Dec–Feb: +0.08
vs 10-yr normal: −8%

Busted

Last winter's polar vortex

“The vortex has a memory.”

Correlation, Nov–Mar: +0.12
Dec–Feb: +0.15
vs 10-yr normal: −5%

Busted

Persistence

“This winter will be like last winter.”

Correlation, Nov–Mar: +0.03
Dec–Feb: +0.08
vs 10-yr normal: −12%

Busted

Cold October

“As goes October, so goes the winter.”

Correlation, Nov–Mar: +0.18
Dec–Feb: +0.19
vs 10-yr normal: −1%

What survives: El Niño strength

Across all winters the link is modest (r = −0.28), and it is not linear. Weak El Niños are as scattered as neutral winters, while the five strongest since 1950 were all mild. In strong El Niño winters, a forecast built on trend plus El Niño strength cut the error against the 10-year normal by about 100 heating degree days, in 98% of bootstrap resamples. Six winters is a small sample, and the 2026/27 El Niño is forecast to be stronger than any of them. The ENSO forecast itself is on /nino/.

Quotable

Ten October folk indicators, 76 winters: none moves US heating demand more than random numbers do.
Siberian snow in October? Its correlation with the winter's gas HDD is +0.14. Random numbers reach that one time in four.
"A cold October means a cold winter" points the right way (r = +0.18) but isn't significant.
"This winter will be like last winter" makes the forecast worse than the trend line alone.
The QBO on its own predicts nothing. Its value is as a filter on the stratosphere, not as a temperature forecast.

How we tested

Target: US heating degree days (65 °F base) from ERA5, weighted by each state's residential and commercial natural gas use (EIA, 2021–25 mean), November–March and December–February, as a departure from the 1951–2026 linear trend. These weights reproduce an industry desk's monthly gas-weighted HDD at r = 0.998. Every indicator uses only what was known by about 1 November: October AO and NAO from ERA5; October Eurasian snow extent and its rate of advance from ERA5 snow depth (40–70°N, 0–180°E); the Gulf of Alaska September–October sea surface temperature; the September–October equatorial wind at 30 and 50 hPa; whether last winter had a major stratospheric warming; last winter's HDD departure; and October's own. El Niño strength is the December–February Niño 3.4 anomaly, which seasonal models forecast well by October. Significance comes from a control of 2,000 random predictors; forecast skill is mean-squared-error skill for winters 1970/71–2025/26, each predicted only from earlier winters. Not tested: the solar cycle.