conus-record-year-2026

Odds of a record-warm 2026 for the contiguous U.S. (data through August 2026)

Run 2026-10-01. Scripts: record_odds.py (models + full-record CV), cv2.py (2000–2025 CV, hinge trend, hindcasts), make_figure.py (pooled estimate + two-panel figure), make_figure_simple.py (single-panel figure). Dependencies in requirements.txt.

Data

• NOAA NCEI Climate at a Glance, contiguous U.S. average temperature, nClimDiv, °F (data/cag_all_months.csv, all months 1895-01 to 2026-08; data/cag_annual.csv 1895–2025; fetch_data.sh refreshes both). URL pattern: .../climate-at-a-glance/national/time-series/110/tavg/all/1/1895-2026/data.csv.
• Checked: NOAA’s annual value equals the unweighted mean of the 12 monthly values (max discrepancy 0.005 °F, rounding). So annual 2026 = (8 × Jan–Aug mean + 4 × Sep–Dec mean) / 12.
• September 2026 is not yet published (NCEI releases ~8th of the month). nClimDiv values can be revised slightly in later releases.

Known values

| Quantity | °F | °C | |—|—|—| | Record annual: 2024 | 55.48 | 13.04 | | Runner-up: 2012 | 55.27 | 12.93 | | 2026 Jan–Aug mean (warmest YTD on record; 2012 was 57.90) | 57.98 | 14.43 | | Sep–Dec mean needed to beat 2024 | 50.50 | 10.28 | | Sep–Dec mean needed to beat 2012 | 49.87 | 9.93 | | Historical Sep–Dec mean, 1895–2025 | 48.66 | 9.26 | | Sep–Dec mean, 2016–2025 | 50.96 | 10.53 | | Record Sep–Dec: 2024 | 52.71 | 11.51 |

Sep–Dec exceeded 50.50 °F in 12 of 131 years, 8 of them since 2000. JJA 2026 (74.37 °F) and August 2026 (75.58 °F) are each the warmest on record.

Models (all predict the Sep–Dec mean; annual follows deterministically)

• A: Sep–Dec ~ year (trend only, baseline).
• B: Sep–Dec ~ Jan–Aug mean + year.
• C: B + August.
• D: B + JJA mean (last three months).
• B2 / D2: B / D with a hinge trend (extra slope after 1980).
• E: AR(1) on standardized monthly anomalies about per-month linear trends, simulated from the observed August anomaly (lag-1 month-to-month correlation ≈ 0.2).
• F: trend fit plus the last 30 years’ detrended Sep–Dec residuals (smoothed bootstrap in the pool). OLS predictive intervals use the t distribution with parameter uncertainty (se_obs). Fit windows: 1895+, 1950+, 1980+.

Cross-validation (rolling origin: predict year t from years < t)

Test years 2000–2025, Sep–Dec mean, °F:

Window	Model	Bias	RMSE	CRPS	90% coverage
1895+	B YTD+trend	+0.76	1.48	0.86	0.73
1895+	D YTD+JJA+trend	+0.64	1.40	0.81	0.73
1950+	D YTD+JJA+trend	+0.51	1.35	0.79	0.69
1950+	D2 hinge	+0.06	1.28	0.74	0.88
1950+	B2 hinge	+0.19	1.27	0.73	0.81
1980+	A trend only	+0.14	1.27	0.73	0.92
1980+	B YTD+trend	+0.13	1.33	0.75	0.88
1980+	D YTD+JJA+trend	+0.01	1.35	0.77	0.92

Bias = observed − predicted. Linear trends fitted from 1895 or 1950 under-predict recent autumns by 0.5–1 °F (CONUS warming has accelerated and the last five Sep–Dec periods are all in the top six), with 90% intervals covering only ~70%. The 1980+ window and hinge-trend fits are unbiased with near-nominal coverage, so the headline pools those. Adding JJA improves CRPS slightly in every window; adding August alone adds little. Detrended correlation of Sep–Dec with Jan–Aug is 0.25, with JJA 0.39, with August 0.21 (1950–2025).

Hindcast sanity check (model D, through August of each year)

| Year | Prior record | YTD rank | P(record), 1950+ / 1980+ | Outcome | |—|—|—|—|—| | 2012 | 54.24 (2006) | 1 | 0.99 / 1.00 | record (55.27) | | 2016 | 55.27 (2012) | 3 | 0.01 / 0.05 | no (54.90) | | 2024 | 55.27 (2012) | 2 | 0.09 / 0.20 | record (55.48), on a record-warm Sep–Dec | | 2025 | 55.48 (2024) | 11 | 0.00 / 0.00 | no (54.62) |

2026 results

| Model | Sep–Dec mean (°F) | SD | P(record) | P(top 2) | |—|—|—|—|—| | D 1980+ | 51.90 | 1.22 | 0.87 | 0.95 | | B 1980+ | 51.88 | 1.23 | 0.87 | 0.95 | | D2 hinge 1950+ | 51.49 | 1.13 | 0.81 | 0.93 | | B2 hinge 1950+ | 51.46 | 1.15 | 0.80 | 0.91 | | E AR(1) 1980+ | 51.28 | 1.18 | 0.75 | 0.88 | | F empirical 1980+ | 51.15 | 1.40 | 0.67 | 0.80 | | Pooled (equal weight) | 51.53 | 1.26 | 0.79 | 0.90 |

Pooled 2026 annual: median 55.83 °F (13.24 °C), 5–95% range 55.12–56.51 °F. Biased long-window linear-trend models (not pooled) give 0.62–0.73 for the record.

Caveats

• The estimate rests on the trend assumption more than on the predictor set: the spread across defensible models is roughly 0.65–0.87, so quote ~80% with “roughly three-in-four to nine-in-ten”.
• Sep–Dec needs to be only moderately warm relative to recent autumns (below the 2016–2025 mean of 50.96 °F). A 2012-style autumn (50.03 °F) would leave 2026 second at 55.33 °F.
• No use of seasonal forecasts (CPC outlook, ENSO state). A strong El Niño winter typically favours a warm northern tier and cool/wet South in Nov–Dec; the net CONUS-mean effect is small and mixed-sign historically, so it was not included.
• Persistence (AR(1)) is weak at monthly scale for CONUS, which is why the autocorrelation model adds little over the regression approaches.

This site is open source. Improve this page.