Predicting the Borana Lunar-Stellar Calendar: An Astronomical Feature Engineering and Machine Learning Approach

Abstract

The Borana calendar of southern Ethiopia and northern Kenya is a unique lunar stellar system where months are defined by new moon conjunctions with specific anchor stars (Triangulum, Pleiades, Aldebaran, Bellatrix, Orion Saiph, Sirius). Unlike arithmetic calendars, it relies on empirical observation by Borana ayyantu (calendar keepers), making prediction challenging. This study aimed to formalize the Borana calendar's astronomical logic using machine learning, predicting new moon conjunction dates, month names (1–12 or intercalary), and day name indices (0–26) from celestial features. Synthetic astronomical data were generated based on synodic month variations, stellar longitudes, and intercalation rules. Features included Moon longitude, angular distance to anchor stars, and cumulative month counts. An LSTM network predicted conjunction dates, while Random Forest classifiers predicted month and day names. Performance was evaluated against baseline arithmetic models. The LSTM achieved Mean Absolute Error of 0.230 days for conjunction dates, improving 7.3% over the mean synodic month baseline. Month classification accuracy reached 94.1%, and day classification 87.5%. Feature importance confirmed angular distance to anchor stars as the strongest predictor. Borana New Year (2027–2070) was predicted between August 18 and October 22. Machine learning successfully captures the Borana calendar's empirical logic, though accurate long term forecasting requires high precision ephemerides and field validation. The framework provides a reproducible methodology for formalizing indigenous timekeeping systems. Future work should integrate JPL ephemerides, ethnographic field data, and open source software tools to support Borana calendar preservation and prediction.