Solar versus Korean lunisolar calendar climatologies for travel destinations: a five-year holdout evaluation at 1,500 locations

1. Introduction

Grouping the same historical weather by different calendars can change climatological estimates. This study compares errors from Gregorian month/day matching (SOLAR) and Korean lunisolar month/day matching (LUNA). LUNA does not mean lunar phase alone. The evaluated systems are one-day climatological estimators, rather than the complete application recommendation systems or short-range numerical forecasts.

2. Data

We included the 1,500 destinations with daily records available in the project. The first 25 years of the 30-year record were used for estimation and the final five for testing. These are ERA5-family reanalysis data, not independent station observations [1]. Daily temperature and precipitation follow the project's local-date aggregation. Cloud evaluation was limited to 274 destinations. Both methods used the same source records and target dates.

3. Methods

SOLAR averaged training samples with the same Gregorian month/day. LUNA matched Korean lunisolar month/day and intercalary-month status exactly [3]. Estimates required at least five valid training samples. All training dates preceded 2016-01-01. The matching window was ±0 days, without interpolation or intercalary substitution. Temperature and precipitation amount used means; rain probability used the frequency of daily precipitation ≥1mm. Clear sky was an exploratory proxy defined as daily mean cloud cover ≤20%, distinct from absence of rain.

4. Evaluation measures

Continuous estimates were assessed with MAE and RMSE; probabilities with the Brier score [2]. Classification accuracy used a forecast probability threshold of 0.5. In the equations, N denotes paired location/date cases, y the verification value, ŷ the estimate, p the forecast probability and o event occurrence (0 or 1). Location/date cases were pooled; countries were not equally weighted.

5. Results

For regular months, SOLAR minimum- and maximum-temperature MAE were 1.9131°C and 2.3390°C, versus 1.9959°C and 2.4165°C for LUNA. LUNA MAE was 4.33% and 3.31% higher, respectively. Precipitation MAE was 0.87% higher and rain Brier score 1.21% higher. The clear-sky proxy had 484,706 regular-month pairs, with LUNA Brier score 0.80% higher. SOLAR had lower pooled errors on all five loss measures in each of the five test years.

6. Intercalary months and seasonal displacement

Intercalary month/day values were never replaced with regular-month values. No intercalary cases met the five-sample training requirement, leaving zero paired comparisons; intercalary performance is therefore undetermined. Seasonal displacement was examined by target Gregorian month and DJF/MAM/JJA/SON, with a separate calendar-distance diagnostic. These labels denote the same calendar-month groups in both hemispheres, not local season names.

7. Regional exploration

LUNA had lower errors at 169 destinations for minimum temperature, 170 for maximum temperature, 481 for precipitation amount and 407 for rain probability. For the clear-sky proxy it prevailed at 100 of 274 destinations. Among country/season/metric combinations with at least ten destinations per year, 32 favoured LUNA in all five years. This was an exploratory search across multiple combinations; the same test period cannot both select a region and independently establish its advantage.

8. Limitations

A single fixed five-year holdout cannot establish future skill. Reanalysis grid and terrain representativeness, destination selection, climate change and missing data may affect the results. Nearby destinations and successive dates are correlated; case counts were not treated as independent trials, and no confidence intervals or significance tests were reported. The ±3/±7-day windows, trip-duration scores, rankings and intercalary performance were outside scope. The clear-sky threshold is exploratory, and this study was not preregistered.

9. Conclusion

SOLAR had lower overall average error than LUNA in this one-day climatology evaluation. Some locations nevertheless had lower LUNA errors; persistence of those regional advantages requires rolling-origin evaluation and independent station records. These findings are limited to the evaluated setting and do not establish a causal lunar effect or weather forecasting for a specific future date.

Data availability and competing interests

Aggregate evaluation results and calendar-displacement diagnostics are provided as supplementary CSV files. Raw weather data and executable code are not distributed on this page. Publisher DILEMMALESS operates the site presenting the compared applications; this report is not an independent third-party evaluation.

References

1. Hersbach H, Bell B, Berrisford P, et al. (2020). The ERA5 global reanalysis. Quarterly Journal of the Royal Meteorological Society, 146, 1999–2049.
2. ECMWF. Forecast User Guide, 12.B: Statistical Concepts — Probabilistic Data.
3. Microsoft. KoreanLunisolarCalendar, System.Globalization.