Network Latency Prediction Intervals Using Adaptive Conformal Inference On Ripe Atlas Measurements

Google Colab notebook and original results (state.pkl) for reproducing the tables and figures reported in the article “Network Latency Prediction Intervals Using Adaptive Conformal Inference on RIPE Atlas Measurements.” The study evaluates 15-minute-ahead round-trip time (RTT) prediction intervals using RIPE Atlas ping measurements to k.root-servers.net from two Jakarta-area probes (AS45725 and AS58821). It compares six point forecasting models: persistence, seasonal naive, ARIMA, LSTM, GRU, and Temporal Convolutional Network (TCN). Prediction interval methods include split conformal prediction, rolling conformal prediction, Adaptive Conformal Inference (ACI), TCN quantile regression (TCN-QR), and conformalized quantile regression with ACI (CQR-ACI). The evaluation covers a main dataset of 7,776 intervals and a routing-change case study of 17,540 intervals, using chronological training, calibration, and test splits. Performance is assessed using MAE, the Diebold–Mariano test, empirical coverage, and Winkler score, with particular emphasis on coverage under traceroute-verified routing changes. Uploading the original state.pkl file to Google Colab regenerates the reported tables and figures without retraining the models. See README.md for instructions and reproducibility details.