A living benchmark for machine learning on streaming data. Every learner reads every stream once, in order, predicting before it trains.

Prequential Single-pass Concept drift CapyMOA River

Per-dataset scores

How to read this board

Elo
The board ordering, computed as in TabArena and Chatbot Arena. On every dataset each pair of algorithms plays one battle, won by the better primary metric (ties are half a win each; every dataset weighs 1). Ratings are the Bradley-Terry maximum-likelihood fit of all battles on the chess scale: 400 points = 10:1 odds of winning a dataset, 100 points ≈ 64%. The task's classic single-tree learner is pinned at 1000, so Elo reads as points above or below a Hoeffding tree. The small range is a 95% interval from 100 bootstrap resamples of the datasets; overlapping ranges mean the order is not settled.
Avg rank
Algorithms are ranked 1st, 2nd, 3rd… on every dataset, then those ranks are averaged. Raw metric values are never averaged across datasets — scales differ.
Wins
Datasets where the algorithm placed first.
Rank per dataset
One cell per dataset (not per time window): teal = ranked 1st on that dataset, clay = ranked last. A solid-teal strip is consistent everywhere; scattered clay shows exactly where it loses. Hover a cell for the dataset name and rank.
Log-loss
The classification ranking metric: mean negative log-probability given to the true label (lower is better, 0 is perfectly confident and correct). A proper scoring rule over predicted probabilities — unlike accuracy-family metrics, it cannot be gamed by echoing recent labels on autocorrelated streams.
AUC (binary)
ROC-AUC over the full stream of predicted scores, shown for binary classification datasets: threshold-free and robust to class imbalance.
κt
Kappa-temporal: skill over the predict-the-previous-label baseline. 100 perfect, 0 no better, negative worse. Display-only; shown as a median — it is unbounded below.
κ
Cohen's kappa: chance-corrected accuracy, robust to class imbalance — the number to trust on skewed streams (imbalanced datasets are flagged in the catalog).
R²
Regression fit, scale-free across datasets: 1 is perfect, 0 is no better than predicting the mean.
ARI
Clustering agreement with the ground-truth labels, scored over evaluation windows: 1 is perfect, 0 is random grouping.
AUC
Anomaly detection ROC-AUC over the full stream of scores: 1 is perfect, 0.5 is random.
Wallclock
Total compute across all datasets and seeds. Read it against rank — in streaming, a small quality gap for a 100× time gap is the whole story.