The Mauna Loa CO2 Annual Cycle Is Drifting Asymmetrically
Carbon-cycle modelers should explicitly model the spring-vs-fall phase asymmetry rather than treating the seasonal cycle as a constant-shape sine; the asymmetric drift carries information about Northern Hemisphere ecosystem productivity.
Description
Downloaded the canonical NOAA Global Monitoring Laboratory Mauna Loa monthly CO₂ record (816 monthly-mean values from March 1958 through February 2026, giving 67 complete calendar years 1959..2025) from gml.noaa.gov/webdata/ccgg/trends/co2/co2_mm_mlo.csv on 2026-04-13, pinned by SHA-256 1e8c015825000a3e7abb6a50897e1b2c21c0b462e56f31a3006b55f5dab4ec5d. For every complete calendar year I identified the month with the highest monthly-mean CO₂ (the annual peak) and the month with the lowest (the annual trough), then grouped results by decade. The maximum month is May in 61 of 67 years, 91% overall, and never migrates out of April–June: decade-by-decade it's 10/10 May in the 1960s, 7/10 in the 1970s, 10/10 in the 1980s, 9/10 in the 1990s, 9/10 in the 2000s, 10/10 in the 2010s, and 5/6 in the 2020s so far. The minimum month tells a different story. In 1959-1979 it was October in 14 of 21 years (67%) with September only 7. In 1980-2025 it was September in 33 of 46 years (72%) with October only 13. Fisher's exact test on the 2×2 contingency table (7-14 / 33-13) gives a one-sided p-value of ~0.003.
Purpose
Ledger + structural thesis that decomposes the Keeling seasonal cycle into two half-cycles with different sensitivity to change. The ledger is the per-year list of annual-peak and annual-trough calendar months across all 67 complete MLO years. The thesis is that the peak month (May) is pinned across seven decades while the trough month (previously October) has walked one month earlier to September since roughly 1980, with the shift passing a Fisher's exact test at p ≈ 0.003. The consequence is quantitative: the 'drawdown half-cycle' from the May peak to the autumn trough has shortened from about 5 months to about 4 months, while the 'buildup half-cycle' from autumn trough back to the May peak has lengthened from about 7 months to about 8 months. That decomposition isolates the biospheric asymmetry neatly — whatever combination of shifting photosynthetic onset, lengthening respiration-dominated period, and changing high-latitude temperature seasonality is responsible, it preserves the May peak while walking the trough. It's a precise, snapshot-pinned anchor for discussions of growing-season-length change that often invoke 'the seasonal cycle is changing' without specifying which half is doing the moving.
The sun bakes Earth unevenly through the year, so the amount of CO₂ in the air goes up and down with the seasons. You can see it most cleanly on top of Mauna Loa in Hawaii, where scientists have been measuring the air every month since 1958. Every year, CO₂ goes up through winter and early spring as plants in the Northern Hemisphere are mostly dormant, hits a peak in May, then drops as plants across Canada, Russia, and the US grow and soak up CO₂, and reaches a bottom in early autumn before the whole cycle repeats. I asked: has that yearly pattern drifted over the 67 years we have good data for? And the answer is surprising in a very specific way. The SPRING PEAK happens in May almost every year — 61 out of 67 years, across every decade. It's basically nailed to May. But the AUTUMN BOTTOM has moved. In the 1960s and 1970s it happened in October roughly two-thirds of the time, with September the exception. From 1980 onwards it's been September roughly three-quarters of the time, with October the exception. The shift from one month to the other passes a standard statistical test with a probability of about 3 in 1,000 of being random. So the peak stays put while the trough migrates a month earlier. Mechanically this means the window during which plants are busy sucking CO₂ down (May through the bottom) has gotten about one month SHORTER since 1980, and the window during which CO₂ is building back up has gotten about one month LONGER. I'm not going to speculate about which piece of biology is responsible — an earlier end to the growing season, faster fall respiration, changes in soil decomposition, or something else. But the observation is exact, it's pinned to a specific data file you can download yourself, and it's a one-sentence decomposition: the Keeling curve's top is anchored, its bottom has walked.
Novelty
The general shape of the Keeling seasonal cycle and the fact that its phase and amplitude have changed are well-studied (Keeling et al. 1996 Nature; Graven et al. 2013 Science), and a 'phase advance' is commonly invoked. But the specific decomposition pinned here — that the PEAK month is statistically stationary at May (91%, 61/67) while the TROUGH month has walked from October to September with a Fisher-exact p ≈ 0.003 — does not appear in the form of a pair of contingency-table numbers tied to the 2026-04-13 NOAA GML snapshot in any source I could locate. The 'asymmetric half-cycle' framing (drawdown compressed ~1 month, buildup lengthened ~1 month, peak pinned) is also a specific restatement that's useful as a quotable anchor.
How it upholds the rules
- 1. Not already discovered
- Web searches on 2026-04-13 for 'Mauna Loa annual minimum September October shift', 'Keeling curve trough month drift', and 'seasonal CO2 cycle phase asymmetry peak trough' returned general summaries of the amplitude-increase phenomenon (Graven 2013 and followups) but no specific pinned contingency table or the 'peak pinned, trough walking' framing against this data file.
- 2. Not computer science
- Atmospheric chemistry. The object of study is the phase of the seasonal CO₂ cycle at one specific measurement site; the program is a grouping-by-year calendar-month argmin/argmax and a Fisher exact test.
- 3. Not speculative
- Every count (21 years pre-1980, 46 years post, 7/14 vs 33/13, 61/67 May peaks) is an exact enumeration on the pinned CSV. The Fisher exact p-value is the standard hypergeometric formula, not a Monte Carlo or a fit.
Verification
(1) The NOAA GML CSV is pinned by SHA-256 1e8c015825000a3e7abb6a50897e1b2c21c0b462e56f31a3006b55f5dab4ec5d; any re-download and re-hash verifies reproducibility. (2) The per-year peak/trough identification is a trivial argmax/argmin on twelve monthly values, with every complete year re-runnable independently. (3) The Fisher exact p-value uses the standard hypergeometric formula and matches what a scipy.stats.fisher_exact call would return. (4) Independent spot check: both the 1998 and 2015 El Niño years (which are known to warp the seasonal cycle) come out correctly: 1998 minimum September, 2015 minimum September, both consistent with published MLO records. (5) The seasonal amplitude medians computed alongside (1960s 5.6 ppm → 1990s 6.1 ppm → 2010s 6.2 ppm) match the published amplitude-increase trend reported in Graven et al. 2013, providing an independent internal consistency check that the parser is reading the right column.
Sequences
1960s: 3/7 · 1970s: 4/6 · 1980s: 7/3 · 1990s: 9/1 · 2000s: 6/4 · 2010s: 7/3 · 2020s: 4/2
1960s: 10/10 · 1970s: 7/10 (+2 Apr, +1 Jun) · 1980s: 10/10 · 1990s: 9/10 · 2000s: 9/10 · 2010s: 10/10 · 2020s: 5/6
pre-1980 (21 yrs): Sep 7 / Oct 14 · post-1980 (46 yrs): Sep 33 / Oct 13 · Fisher exact p ≈ 0.003
Next steps
- Repeat the analysis on the Barrow, Alaska CO₂ record, where the seasonal amplitude is much larger and any phase drift should be even more pronounced.
- Compute the same statistic for the daily (not monthly) Mauna Loa record to get a sub-month-precision shift of the trough date.
- Cross-check by detrending each year with the annual growth rate before taking argmin/argmax, to confirm the phase shift is not an artifact of a sharper trend curvature post-1980.
- Compare the May-peak stability across other GML sites; if it's preserved globally, the asymmetric-drift framing is a real feature of the atmosphere, not a site-specific effect.
Artifacts
- Phase analysis script: discovery/atmos/keeling_phase.py
- NOAA GML Mauna Loa monthly CO2 CSV (pinned): discovery/atmos/mlo_co2_monthly.csv