Predicting Internet Outage Duration from Cause, Type, and Scope: Linear Regression Tested on Unseen Countries

Internet outages disrupt education, commerce and communication, yet little is known about how well simple, publicly available descriptors of an outage predict how long it will last. This study asked whether an outage’s assigned cause, type and scope (number of affected networks, locations and countries) can predict its duration using linear regression, on events from countries the model has not seen, and whether prediction accuracy differs by cause. The data were 862 outage records from the Cloudflare Radar Outage Center (January 2022 to August 2026; 137 countries). Duration (natural logarithm of hours) was modeled with ordinary least squares. Models were tested by grouped cross-validation, in which all records from a country were held out together (5 folds, repeated 20 times), and compared with a no-predictor baseline, a cause-only model, gradient boosting and a random forest. A conventional random-row split was run for comparison. For unseen countries, the full linear model explained little of the variation in duration (R^2 = 0.081, SD = 0.055 across repeats; 95% bootstrap interval −0.003 to 0.148), compared with −0.067 for guessing the average; typical predictions were off by a factor of about 3.3. Accuracy differed by cause: errors were smallest for unknown-cause, technical-problem and maintenance outages (typical factor 2.1–2.5) and largest for military action and other causes (6.1–6.9). Evaluation design mattered more than model choice: the same linear model scored R^2 = 0.257 when rows were split at random, because repeated, near-identical events such as scheduled exam shutdowns in Iraq appear on both sides of the split (data leakage), and gradient boosting fell from 0.328 to −0.094 when countries were held out. In an exploratory analysis, government-directed outages whose descriptions mention exams had a median duration of 2.0 h (n = 239), compared with 8.3 h for other government-directed outages (n = 185), and adding simple description flags raised R^2 to 0.229; because these rules were chosen after inspecting the data, this result needs confirmation on new events. Limitations include the concentration of records in a few countries (Iraq alone contributes 20%), cause labels that are inferred rather than confirmed, and the lack of a test on events recorded after the data cutoff. Cause, type and scope therefore carry limited information about outage duration, and studies that include repeated events should hold out whole groups when testing models.