Potential periodic signals in blazars: significance, forecasting and deep learning

Abstract

Blazars exhibit variable emission on diverse timescales. Some light curves show signs of quasiperiodic oscillations (QPOs), which may encode clues regarding the physical processes behind the emission or point to supermassive binary black holes. We analyzed five blazars with previously reported high significance year-long QPOs, applying the Lomb-Scargle periodogram and Weighted Wavelet Z-transform methods to Fermi-LAT data up to early 2025. We furthermore examined an additional source (PKS 0139-09), where nascent QPO may be present. As the light curves showed longer term trends, we detrended the data using an STL decomposition, which often revealed a large seasonal component. We find that detrending generally leads to an increase in the strength of the QPO signal. However, except for PG 1553+113, where a clearly persistent QPO signal is present, we detect transience on a timescale of $\lesssim$4000 days. We then forecast the light curves over the following four years, using a traditional statistical method as well as a Transformer-based deep learning model. Applied to a test set, the latter showed significant success in predicting behavior that seems unexpected from simple inspection of the past data. Analyzing the extended time series suggests a markedly weaker QPO signals over the coming years in cases where the transient behavior appears near the end of the observational data. In contrast, in the nascent candidate QPO source (PKS 0139-09) the signal is expected to strengthen significantly. These predictions, which may reflect the physical origin of the QPOs, can be tested against future data.