Cosmology & Space (🚀🌠) Deep ResearchDeep

solar-activity-across-centuries-sunspot-flare-analysis

An extensive statistical analysis of more than 300 years of sunspot records, historical solar activity reconstructions, and modern NOAA GOES solar flare observations. This research examines long-term…

solar-activity-across-centuries-sunspot-flare-analysis
Quick Read

More than 300 years of sunspot records and modern GOES flare reports reveal predictable cycle structure alongside sharp, irregular flare extremes.

Solar Cycle 25 reached a smoothed peak of 160.9 in October 2024, 38% above Cycle 24. Across matching January–November windows, the M-or-stronger flare count declined from 834 in 2024 to 246 in 2026. The study tests outliers, regression relationships, and historical backtests, then forecasts sunspot and strong-flare activity into 2027.

Forecast boundary: November 2026 observations and projections are a dated snapshot through November 21 UTC. The forecast ranges are probabilities, not guarantees; this analysis does not establish solar causation of terrestrial weather.

THE CYCLE

11.0 years · Dominant period in monthly sunspot records

THE PEAK

160.9 · Solar Cycle 25 smoothed peak in October 2024

THE RELATIONSHIP

3.93× · Fitted M+ flare rate at 100 vs. 50 sunspots

THE FORECAST

14.8 · Expected strong flares over the subsequent 30 days

01 · SOLAR RESEARCH

The current cycle in context

Cycle 25’s smoothed peak in October 2024 was 38% higher than Cycle 24’s. By October 2026 the monthly sunspot number had fallen to 45.7, down 60% from October 2025. The first 21 November days averaged 90.0, which materially raised the November estimate.

In the GOES science catalog, M-or-stronger flares fell from 834 in the January 1–November 19 window of 2024 to 246 in the matching 2026 window, a decline of 70.5%. The same windows contain 49 and 13 X-class events.

Figure 12. Calendar-matched GOES counts show the recent decline in M-or-stronger and X-class events.

GOES events from January 1 through November 19

Year M or stronger X class Calendar days
2023 340 11 323
2024 834 49 324
2025 314 16 323
2026 246 13 323

02 · SOLAR RESEARCH

Sunspots, cycle shape and deviations

Figure 1. Annual means and centered 13-month means establish the historical range. The lower panel shows the two latest cycles.

The highest completed-cycle peak was 285.0 in March 1958. Complete cycles lasted 9.0–13.6 years, with a median of 11.1 years. The 24 peak values exhibit substantial multi-decade variation. A linear trend was +4.6 sunspot-number units per century (p = 0.793).

Figure 2. The horizontal lines show the two sample means at the best scanned split. Cycle 25 remains open and is excluded from the test.

The strongest one-break candidate falls at Cycle 17, starting in 1933. Average peaks on either side were 168.0 and 200.1. The search-adjusted p-value of 0.839 provides weak evidence for a single persistent shift in the mean peak.

Selected solar cycles

Cycle Minimum Peak month Peak Rise (years) Duration (years)
3 Jun 1775 May 1778 264.3 2.9 9.2
5 Apr 1798 Feb 1805 82.0 6.8 12.2
6 Jul 1810 May 1816 81.2 5.8 12.8
19 Apr 1954 Mar 1958 285.0 3.9 10.5
22 Sep 1986 Nov 1989 212.5 3.2 9.9
23 Aug 1996 Nov 2001 180.3 5.2 12.3
24 Dec 2008 Apr 2014 116.4 5.3 11.0
25 Dec 2019 Oct 2024 160.9 4.8 Ongoing

Figure 3. Historical cycle shapes and the fitted relationship between rise time and peak intensity. The shaded regression band covers the fitted mean.

Across 24 completed cycles, each additional year from minimum to peak was associated with a 38.2-point lower peak (95% confidence interval 26.1–50.2, R² 0.56). This is the observed rise-time relationship often called the Waldmeier effect.

Figure 4. A strong period near 11 years dominates monthly counts. After removing an 81-day trend, the daily series shows a weaker secondary recurrence near 26 days.

The daily residual correlation at its strongest 22–33 day lag was only 0.09 at 26 days. That gives a repeatable timing signal with limited day-to-day predictive power.

Figure 5. The local score identifies unusual days relative to nearby conditions. The rankings show the largest raw counts and the sharpest 27-day changes.

The highest raw daily number was 528 on Aug 26, 1870. Its historical SILSO row lists 1 observation. In the modern local comparison, Jan 5, 2013 reached 162 against a local median of 60. The largest negative local score was -3.22, below the chosen threshold in absolute magnitude.

Largest separated daily departures since 1981

Date Direction Observed Local median Difference Local score
Jan 5, 2013 Above local median 162 60 +102 +5.29
Oct 22, 1985 Above local median 85 10 +75 +4.60
Mar 27, 2001 Above local median 352 139 +213 +4.56
Nov 29, 2020 Above local median 96 24 +72 +4.41
Mar 4, 1991 Below local median 76 224 -148 -3.22
Nov 26, 2013 Below local median 37 124 -87 -3.17
Jan 27, 2004 Below local median 0 68 -68 -2.96
Dec 6, 2000 Below local median 78 157 -79 -2.88

Figure 6. The hemisphere supplying most spots changes with time. At the 2024 cycle peak the southern contribution was substantially larger.

The southern hemisphere supplied 65.6% of the summed 2024 hemispheric sunspot number. In 2025 the northern share returned to 51.1%. The 13-month hemispheric maxima in Cycle 25 appeared in April 2023 in the north and October 2024 in the south.

Figure 7. Monthly spot number and corrected area are strongly associated. The right panel shows the empirical fit on the original measurement scales.

The log-transformed area regression explained 94.3% of the variation across 1,782 paired months. The largest daily corrected total area in the Mandal composite was 8,596 millionths of the solar hemisphere on Oct 29, 2003.

03 · SOLAR RESEARCH

Flare extremes and the sunspot relationship

Figure 8. The optical flare index and satellite counts show similar cycle-scale variation in their own units. The shaded early X-ray segment uses legacy reports.

The H-alpha–based index reached its annual high of 19.42 in 1957. Its early Ondřejov segment and later Kandilli segment are shown with their source boundary. The NOAA analysis below uses 82,231 events in the recalibrated science catalog, including 6,072 M-or-stronger and 366 X-class events.

Figure 9. Flux saturation and event frequency are shown separately. The four hatched measurements preserve their lower-bound status.

Nov 4, 2003 registered at least X24.8 in the science catalog. The largest unsaturated peak was X20.6 on Apr 15, 2001. The busiest UTC day was Dec 29, 2024, with 21 M-or-stronger events.

Figure 10. The negative binomial regression estimates the rate at each monthly sunspot number. Residual spikes identify strong-flare months poorly explained by spot count alone.

At a monthly sunspot number of 100, the fitted M-or-stronger rate is 3.93 times the rate at 50 (95% interval 3.16–4.90). Allowing separate baseline rates for Cycles 23–25 gives 3.83 times (95% interval 2.98–4.92). The cycle-specific effects at a fixed sunspot number had wide intervals.

The strongest positive monthly departure, September 2005, logged 57 events against a fitted mean of 3.4. February 2026 had 85 against 11.4 even as the broad cycle was declining.

Strong flare months above the fitted rate

Month Sunspot mean Observed M+ Fitted M+ Residual score
Sep 2005 37.2 57 3.4 +14.15
Sep 2017 43.7 44 4.3 +8.43
Jan 2005 48.1 47 5.2 +7.49
Feb 2026 77.4 85 11.4 +6.34
Mar 2015 54.5 43 6.4 +5.38
Jun 2007 21.3 12 1.7 +4.83

Figure 11. The independent STIX list and Earth-view GOES series show similar timing at monthly scale. Instrument coverage and event thresholds differ.

04 · SOLAR RESEARCH

Forecasts and backtests

November sunspots and the next 12 months

The October-only model projected 46.0 for November. Using days 1–21 lifts the November estimate to 80.2 (80% range 69.4–95.6). The selected day-21 method had a historical holdout mean absolute error of 6.9 sunspot-number units. Its historical inputs use retrospectively released SILSO daily data.

Figure 13. The orange point estimates the full November mean with the first 21 days in hand. It initializes the purple forecasts. The lower panel compares the original monthly forecasts on historical holdouts.

The cycle analogs model scored 24.5 mean absolute error at 12 months versus 34.2 for the last-value baseline. At one month its error was 18.6, with 16.3 for the ridge autoregression. Model selection used pre-1980 forecasts. The partial-month update has its own calibration and holdout evaluation.

November nowcast and subsequent monthly forecasts

Month Basis Sunspot number 80% range 95% range
Nov 2026 Partial-month estimate 80.2 69.4–95.6 61.5–107.2
Dec 2026 Forecast 46.2 3.7–80.7 0.0–108.9
Jan 2027 Forecast 47.9 6.2–92.1 0.0–137.0
Feb 2027 Forecast 51.0 8.3–86.1 0.0–99.9
Mar 2027 Forecast 50.5 0.0–82.1 0.0–104.5
Apr 2027 Forecast 56.4 7.3–102.5 0.0–122.8
May 2027 Forecast 53.5 8.7–89.1 0.0–106.0
Jun 2027 Forecast 37.8 0.0–79.3 0.0–104.9
Jul 2027 Forecast 42.3 2.9–79.6 0.0–115.3
Aug 2027 Forecast 46.1 0.0–77.5 0.0–105.3
Sep 2027 Forecast 43.8 0.0–79.5 0.0–88.1
Oct 2027 Forecast 39.6 0.0–76.5 0.0–124.4
Nov 2027 Forecast 41.1 0.0–81.1 0.0–104.9

Sunspot model mean absolute error on the original monthly holdout

Model 1 month 3 months 6 months 12 months
6-month mean 18.0 20.6 25.7 37.6
Cycle analogs 18.6 18.6 20.1 24.5
Damped trend 17.5 19.8 24.4 31.6
Persistence 16.6 20.0 23.8 34.2
Ridge autoregression 16.3 18.7 22.2 29.0

Probability of a strong flare day

The November 21 archive implies an 8.3% probability of at least one M-or-stronger flare on November 22 UTC. The model combines lagged daily sunspot counts with recent flare activity. The held-out Brier score was 0.095 across 3,287 days, 18% lower than the recent-rate baseline and 45% lower than the expanding climatology. A year-block bootstrap puts the latter skill estimate at 31% to 60%.

Figure 14. Held-out scores and calibration group days with similar probabilities. The bars around observed frequencies show binomial sampling uncertainty.

Holdout score for predicting at least one M-or-stronger flare

Model Brier score ROC AUC Average precision
Sunspots + flares logistic 0.0947 0.914 0.760
Gradient boosting 0.0954 0.914 0.750
Sunspot logistic 0.1095 0.892 0.670
Trailing 27 days 0.1154 0.872 0.611
Climatology 0.1727 0.425 0.222

Strong flare counts over one week and one month

From the November 21 snapshot, the count models estimate 3.4 M-or-stronger events over the following seven days and 14.8 over 30 days. The respective 80% ranges are 0–7 and 5–37.

Figure 15. The forecast points are expected counts, so they can be fractional. The colored lines show the empirical forecast ranges.

Strong flare count forecasts

UTC forecast window Expected count 80% range 95% range
Nov 22 – Nov 28 3.4 0–7 0–22
Nov 22 – Dec 21 14.8 5–37 2–55

The far tails are uncertain. The nominal 95% 30-day flare range covered 83.3% of historical holdout windows. Its nominal 80% range covered 76.5%. The seven-day 80% range covered 80.1%.

Figure 16. Observed coverage for the updated sunspot and flare count models. Dashed lines show the two nominal coverage levels.

05 · SOLAR RESEARCH

Early observations and isotope evidence

Figure 17. The oldest plotted series are isotope reconstructions. Telescope-era comparisons show how independent group reconstructions differ through sparse periods.

Decadal radiocarbon estimates in the archive extend back roughly 11,400 years before 1950. The Wu series combines radiocarbon and beryllium evidence. Historical group-count databases begin in 1610 and have sharply uneven observing coverage. Dated naked-eye sightings and early optical flares provide individual records. The modern forecast models are calibrated using the consistent SILSO and GOES series.

06 · METHODS & VALIDATION

Methods and validation

Observed changes and unusual events

  • Sunspot cycle peaks use the published centered 13-month SILSO series between published minima. Cycle 25 is included descriptively and excluded from completed-cycle fits. The single-break scan searches splits with at least five cycles per side. Its p-value compares the largest scanned contrast with 20,000 Gaussian AR(1) simulations at the estimated lag-one correlation.
  • Daily outliers use the difference from an 81-day centered median divided by the larger of 15 or the local interquartile range divided by 1.349. Flagging requires an absolute score of 3.5. The scale is robust and descriptive. It has no standard-normal probability interpretation. The ranked independent dates are separated by more than 14 days.
  • Flares are assigned to the UTC date of peak irradiance. A strong flare has a peak of at least 10⁻⁵ W/m², and X class begins at 10⁻⁴ W/m². The catalog’s zero days represent zero reported events. Continuous instrument uptime would require minute-resolution flux coverage. The NOAA science archive supplies the historical fit. The SWPC snapshot supplies the final two days of predictive inputs.
  • The monthly rate model uses the log of 1 + sunspot number / 50, with an exposure offset for the number of calendar days. Negative binomial dispersion accommodates clustering. Confidence intervals use a Bartlett HAC covariance with 12 monthly lags. The monthly residual table ranks upper-tail probability from the fitted retrospective model. Peaks and month-level fits do not imply an attribution to individual active regions.

Forecast validation

  • Monthly sunspot models compare persistence, a trailing mean, damped trend, ridge autoregression and completed-cycle analogs. Development labels finish in 1979. Holdout origins begin in 1980 and targets finish by 2025. Cycle-analog candidates require an already completed comparison cycle and a six-month buffer for minimum recognition.
  • The partial-month comparison evaluates four day-21 methods on 1950–1979 development and 1981–2025 holdouts. In the updated monthly model, the estimated last-month mean is substituted at each retrospective origin. Empirical error intervals are calibrated on 1981–2004 targets and checked on 2005–2025 holdouts.
  • Day and count models use information through the prior UTC day. The day models compare climatology, a 27-day rate, logistic regressions and gradient boosting. Expanding annual fits are scored on 2005–2016 development and 2017–2025 holdouts. Each held-out calendar year enters training only after its predictions are issued. The historical runs use revised records as retrospective inputs.
  • Seven- and 30-day count models predict the sum within each window, using nonoverlapping evaluation windows for each horizon. Candidate models use a climatological mean, the prior 27-day rate or a negative binomial regression. Earlier out-of-sample residuals provide count ranges scaled by the square root of 1 + forecast. Their observed coverage is reported above.

SOURCES

Primary data and sources

The companion charts and models ZIP contains 17 individual figures, analysis tables, forecast scores, and the code used to recreate them.

Research archive: November 22, 2026. Observation cutoff: November 21 UTC. October 2026 sunspot data are provisional. Solar Cycle 25 remains open; model forecast intervals require calibration and are not deterministic event predictions.
Pattern Nexus Research Note

This study reports solar-cycle structure, observed flare outliers, fitted historical relationships, and backtested forecasts. It does not claim that sunspot counts alone explain every extreme flare or establish a causal connection to polar-vortex dynamics.

Article overview

An extensive statistical analysis of more than 300 years of sunspot records, historical solar activity reconstructions, and modern NOAA GOES solar flare observations. This research examines long-term solar cycles, extreme outliers, deviations from historical patterns, changing relationships between sunspots and powerful flares, and evidence of structural shifts in solar behavior. Using regression analysis, spectral analysis, statistical modeling, and out-of-sample forecasting, the study evaluates Solar Cycle 25 against centuries of historical data and develops probabilistic projections of future solar activity. Includes detailed scientific charts, historical comparisons, anomaly detection, and predictive model performance.

Downloads (1)

What's Your Reaction?

Like Like 0
Dislike Dislike 0
Love Love 0
Funny Funny 0
Wow Wow 0
Sad Sad 0
Angry Angry 0
Nexus (Christopher)
Research by

Nexus (Christopher)

Founder of Pattern Nexus. I research markets, macro, geopolitics, AI, history, ancient systems, and the patterns most people overlook. I’m also building Market Radar, a trading scanner designed to read pressure, risk, confirmation, and setup quality before ch…

Comments (0)

User