In February at Abisko, 7.8 in 10 four-night stays had a clear, dark night with the aurora active

Method in five lines

1. A night runs local noon to noon (Europe/Stockholm); dark means the sun more than 12 degrees below the horizon at 68.35 N, 18.83 E.
2. Each hourly SMHI cloud reading at Abisko Aut (station 188790) must be 25% or less; 113 (sky obscured) counts as cloudy.
3. Each hour takes the Kp of its 3-hour GFZ block, which must be 2 or more; Abisko sits under the aurora oval, so Kp 2 is a low bar for activity, not a guarantee of a show.
4. A night is good with at least 2 such hours; a trip window of k nights is a success if any of its nights is good, over every start date in September to April.
5. A night with any missing dark-period hour is excluded, and so is any window containing it. Gaps are never filled.

What this table can and cannot say

• Cloud is the Abisko Aut station reading, in the Abisko valley. A trip that drives or walks to clearer sky may do better; one that stays in cloud will not.
• Kp is a planet-wide index. Kp 2 or more marks a night with some geomagnetic activity. It is not a sighting record, and we do not claim how bright the aurora was.
• Seven winters is a short record, and they run from the minimum to near the peak of the 11-year solar cycle. Geomagnetic activity tends to be higher near the peak. The by-winter file shows the early winters well below the later ones. A quieter stretch of the cycle would likely give lower figures. This is the expectation, not something this data shows.
• Past windows describe the past. They are not a forecast for any given trip.

Independent check

Independent check by a reviewer who read only the method (spec.md and the method section above), not odds.py. Run on the raw files in data/raw/ (SMHI 188790 cloud, GFZ Kp).

Different sun method. Astronomical Almanac low-precision sun position (not the NOAA-style formula), dark = sun below -12 degrees. Reference: sunrise-sunset.org API (Abisko 68.35 N, 18.83 E, one request per date, UTC), nautical twilight (= -12 degrees). timeanddate.com answered 403 to a plain fetch and was not pursued or routed round.

Within about 2 minutes (mine is found to the whole minute). That is well inside what could move an hourly count.

Cell-by-cell result. 100 cells compared: the full main table (8 months x 5 lengths, 40 cells) and the February and December rows of every sensitivity variant (60 cells). Maximum absolute difference 0.6 percentage points. No cell differs by more than 2 points. Every window count (n) matches exactly, and usable nights per month match the table above exactly. The only visible differences are one-night shifts in a handful of cells (for example Sep 1 night: 21.1% mine vs 20.6% builder, one night in 180, from a marginal twilight hour on the 12-degree line).

Other checks.

1. The three plain-English sentences are true to the table: Feb 4 nights 78.5% (7.8 in 10); winter range for that cell 56% (2019-20) to 96% (2025-26); Dec 3 nights 58.7% (5.9 in 10) and 1 night 33.0% (3.3 in 10); Apr 5 nights 24.3% (2.4 in 10).
2. The "about 80% over three nights" comparison is fair. Recomputed: Oct to Mar three-night windows 63.7% (1211 windows), monthly 53% to 70%; loosest definition 92.7%; cloud-only loosening 70.7%; Kp-only loosening 79.8%. It states the figure is the operator's, undefined to us, and labels the comparison unverified. It does not call the operator wrong.
3. The solar-cycle caveat is worded as an expectation ("tends to be", "would likely", "the expectation, not something this data shows"). One wording note: "they fall around the peak" is loose. 2019-20 sits near the cycle minimum, and the winters run from minimum to near maximum. Suggest "they run from the minimum to near the peak of the 11-year solar cycle". The by-winter spread (Feb 4 nights 56% rising to 96%) supports the point but is not proof of cause, and the text already says so.
4. Credits: SMHI (CC BY 4.0 SE, modified) and GFZ (CC BY 4.0, doi:10.5880/Kp.0001, modified) are present. No sunspot data is used: the GFZ files hold only Kp, datetime, status and meta.

Verdict: CONFIRMED WITH NOTES. The numbers reproduce. The one note is the cycle-phase wording in item 3. Limits of this check: it re-implements the same method definition (a wrong definition would repeat in both), and it uses the same two raw files, so it does not test the downloads themselves.

Usable nights per month

Method in five lines

1. A night runs local noon to noon (Europe/Oslo); dark means the sun more than 12 degrees below the horizon at 69.65 N, 18.96 E (Tromsø).
2. Each hourly MET Norway cloud reading at Tromsø-Langnes airport, station SN90490, must be 25% or less (2 oktas or fewer); the source reports oktas, so 1 okta = 12.5% and 2 = 25%; a reading of 9 (sky obscured) counts as cloudy.
3. Each hour takes the Kp of its 3-hour GFZ block, which must be 2 or more; Kp 2 is a low bar for activity, not a guarantee of a show.
4. A night is good with at least 2 such hours; a trip window of k nights is a success if any of its nights is good, over every start date in September to April.
5. A night with any missing dark-period hour is excluded, and so is any window containing it. Gaps are never filled.

The station

Cloud comes from Tromsø-Langnes airport, MET Norway station SN90490 (69.6767 N, 18.9133 E), about 3 km from Tromsø. It is the airport, on the same island as the town. The weather station in the town itself (SN90450) reports total cloud only every 6 hours, which is too sparse for an hourly method, so it was not used. Langnes has reported hourly cloud since 2007. About 8% of its hourly readings are 9 (sky obscured), which counts as cloudy. Winter rule: a winter is used only if at least 80% of its September to April nights are usable (no missing dark-period hour). All ten passed; the lowest, 2017-18, had 91.7% usable.

What this table can and cannot say

• Cloud is one station's reading, 3 km from Tromsø. Tromsø sits on an island among fjords and mountains, and the airport is on the same island, so cloud on a trip that drives inland or north can differ from the airport's in either direction. This table does not measure that.
• Kp is a planet-wide index. Kp 2 or more marks a night with some geomagnetic activity. It is not a sighting record, and we do not claim how bright the aurora was.
• Ten winters is a short record, and they run from the declining years of the last 11-year solar cycle, through its minimum around 2019, to near the peak of the current one. Geomagnetic activity tends to be higher near the peak. For three-night windows from October to March, the by-winter file has 2018-19 lowest at 49% and 2025-26 highest at 76%. A quieter stretch of the cycle would likely give lower figures. This is the expectation, not something this data shows.
• At about 70 N, September and April are the months most sensitive to the sun formula: the sun sits only just below the 12 degree line for part of those nights, so a few minutes of error in the dark-hours count can move a night in or out. The independent check for Tromsø (below) moved September by up to about 3 points and February by up to about 1.3, all traced to nights where the sun sits within 0.2 degrees of the 12 degree line. Read September and April as soft, and the main table as accurate to a few points.
• Past windows describe the past. They are not a forecast for any given trip.

Independent check

The check was done by an independent reviewer who read only the method, not odds.py. They wrote their own code, used the Astronomical Almanac sun position, and ran it on the same raw files.

• Main table: the largest difference was 3.1 points, in September. February moved by 0.8 to 1.3 points and every other month by 0.5 points or less.
• Sensitivity table (Feb and Dec rows): the largest difference was 2.2 points.
• Why the figures differ: every difference traces to a night when the sun was within 0.05 to 0.17 degrees of the 12 degree line.
• Window counts: all match except March, where usable nights came to 302 against 301.
• Single-winter range for Feb 4 nights: 61% to 100% here and 50% to 100% in the check. The 2022-23 winter turns on one borderline night, so the sentence now says "roughly half to all".
• Alta: the blocked reason was confirmed from Frost's own series list.
• Client ID: not present in any file.

Verdict: CONFIRMED WITH NOTES. Both notes have been applied.

Usable nights per month

Method in five lines

1. A night runs local noon to noon (Europe/Helsinki); dark means the sun more than 12 degrees below the horizon at 66.5 N, 25.73 E (Rovaniemi).
2. Each hourly FMI cloud reading at Rovaniemi railway station, station 101928, must be 25% or less (2 oktas or fewer); the source reports oktas, so 1 okta = 12.5% and 2 = 25%; a reading of 9 (sky obscured) counts as cloudy.
3. Each hour takes the Kp of its 3-hour GFZ block, which must be 2 or more; Kp 2 is a low bar for activity, not a guarantee of a show.
4. A night is good with at least 2 such hours; a trip window of k nights is a success if any of its nights is good, over every start date in September to April.
5. A night with any missing dark-period hour is excluded, and so is any window containing it. Gaps are never filled.

The station

Cloud comes from Rovaniemi railway station, FMI station 101928 (66.4983 N, 25.7088 E), about 1 km from Rovaniemi. It is the FMI station in the town itself, and the nearest to the town centre with an hourly cloud record. The airport's automatic station (137190), 8 km from the centre, was the first choice, but it has almost no hourly cloud for the winters 2017-18 and 2018-19, so it could not give a full run of winters. As a cross-check, over the 8 winters where both stations have a usable record, three-night windows from October to March had a good night 47.1% of the time on this station's cloud and about 47% on the airport's (47.5% here, 47.2% in the independent check) (1,368 and 1,394 windows), so the choice of station does not move the headline.

What this table can and cannot say

• Cloud is one station's reading, 1 km from Rovaniemi. Town-centre light and a river valley are not measured here; a trip that leaves town for darker sky may see different cloud.
• Kp is a planet-wide index. Kp 2 or more marks a night with some geomagnetic activity. It is not a sighting record, and we do not claim how bright the aurora was.
• Ten winters is a short record, and they run from the declining years of the last 11-year solar cycle, through its minimum around 2019, to near the peak of the current one. Geomagnetic activity tends to be higher near the peak. For three-night windows from October to March, the by-winter file has 2020-21 lowest at 31% and 2024-25 highest at 74%. A quieter stretch of the cycle would likely give lower figures. This is the expectation, not something this data shows.
• Past windows describe the past. They are not a forecast for any given trip.

Independent check

An independent reviewer read only the method (not odds.py). They used their own code with the Astronomical Almanac sun position and the same raw files.

• Window counts and usable nights match on every cell.
• In October to March the cells agree closely. In September and April they differ by up to 4.7 points, because of how sensitive the sun formula is at 66.5 N. For example, April 5 nights is 41.0% here and 36.3% in the check.
• Twilight times are within 4 minutes of sunrise-sunset.org on two dates.
• The plain-English sentences and the single-winter range (41% to 82%) are confirmed.
• Verdict: CONFIRMED WITH NOTES. The notes have been applied: the April sentence is now rounded, the shoulder-month caveat is added, and the winter rule is stated.

Usable nights per month

Method in five lines

1. A night runs local noon to noon (Europe/Helsinki); dark means the sun more than 12 degrees below the horizon at 68.42 N, 27.41 E (Saariselkä).
2. Each hourly FMI cloud reading at Ivalo airport, station 102033, must be 25% or less (2 oktas or fewer); the source reports oktas, so 1 okta = 12.5% and 2 = 25%; a reading of 9 (sky obscured) counts as cloudy.
3. Each hour takes the Kp of its 3-hour GFZ block, which must be 2 or more; Kp 2 is a low bar for activity, not a guarantee of a show.
4. A night is good with at least 2 such hours; a trip window of k nights is a success if any of its nights is good, over every start date in September to April.
5. A night with any missing dark-period hour is excluded, and so is any window containing it. Gaps are never filled.

The station

Cloud comes from Ivalo airport, FMI station 102033 (68.6134 N, 27.4189 E), about 22 km from Saariselkä. Saariselkä has no FMI station reporting hourly cloud: the two stations at the resort (the tourist centre and Kaunispää) returned no cloud values in the checks we ran. Ivalo airport is the nearest FMI station that does, and it is to the north of the resort.

What this table can and cannot say

• Cloud is one station's reading, 22 km from Saariselkä. Saariselkä sits among fells and Ivalo airport does not, so cloud at the resort can differ from the airport's in either direction. This table does not measure that.
• Kp is a planet-wide index. Kp 2 or more marks a night with some geomagnetic activity. It is not a sighting record, and we do not claim how bright the aurora was.
• Ten winters is a short record, and they run from the declining years of the last 11-year solar cycle, through its minimum around 2019, to near the peak of the current one. Geomagnetic activity tends to be higher near the peak. For three-night windows from October to March, the by-winter file has 2020-21 lowest at 30% and 2024-25 highest at 72%. A quieter stretch of the cycle would likely give lower figures. This is the expectation, not something this data shows.
• Past windows describe the past. They are not a forecast for any given trip.

Independent check

An independent reviewer read only the method (not odds.py). They used their own code with the Astronomical Almanac sun position and the same raw files.

• The largest difference anywhere is 1.7 points in the main table and 1.0 points in the sensitivity rows.
• Window counts and usable nights match on every cell.
• The plain-English sentences and the single-winter range (54% to 82%) are confirmed.
• Verdict: CONFIRMED WITH NOTES. The one note was to state the winter-inclusion rule, and that is now done.

Usable nights per month