Answers are grounded in this paper and cite their sources.
Weekly cycles in cyber-threat activity are well documented, but how their timing varies with national weekends is not. Across 384 series from 192 countries, the weekly cycle is nearly universal. In the 43 countries whose weekend we individually sourced, its phase is statistically associated with the local working week. Under the dataset’s date labels, activity is lowest on each country’s own first working day (Sunday) in every Friday–Saturday-weekend country. In just over half of Saturday–Sunday countries, it is lowest one day later (Tuesday rather than Monday). This separation survives any uniform shift of the labels; the absolute weekday depends on an undocumented source timezone. Amplitude is larger under Friday–Saturday weekends (p = 0.0010). Auditing the benchmark, we also found that its ambiguous date format, read with a common parsing idiom, silently discards 59.9% of the data. A full factorial of day-of-week specification, multi-step strategy, and learner, with reimplemented baselines, rolling-origin evaluation, and Diebold–Mariano/Holm inference, shows that specification improves every stable pairing at H = 10 (factors 0.57–0.87) and wins on 190 of 192 countries. It is not sufficient: the multilayer perceptron still loses to a seven-observation moving average. No learned model significantly beats a fixed-window seasonal decomposition at either operational horizon. Accuracy here is governed by specification, not architecture.
No takes yet. Share an insight, caveat, or question.
Alwehaibi et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: