Skip to content

Yahoo S5 — application guide (no redistribution)

What is this data? Production traffic metrics from Yahoo services (A1: real, A2–A4: synthetic with controlled trend/seasonality/changepoints), univariate, hourly. A classic web-metrics benchmark — access requires a research application.

Yahoo S5 may not be redistributed, so this repository provides no download script — only a loader for locally placed files (CLAUDE.md §2).

How to apply

  1. Visit https://webscope.sandbox.yahoo.com/catalog.php?datatype=s ("S5 - A Labeled Anomaly Detection Dataset")
  2. Sign in with a Yahoo account and submit the research-purpose request form (an academic email address is recommended)
  3. Download ydata-labeled-time-series-anomalies-v1_0.tgz from the approval email

Local placement

data/yahoo_s5/
├── A1Benchmark/real_1.csv ...      # real traffic (67 series)
├── A2Benchmark/synthetic_1.csv ... # synthetic (100)
├── A3Benchmark/ ...                # synthetic + trend/seasonality (100)
└── A4Benchmark/ ...                # synthetic + changepoints (100)

Loader: load_yahoo(benchmark="A1Benchmark", series="real_1.csv")

Known flaws

  • A2–A4 are synthetic — do not use them to support claims about real-world generalization.
  • There is no official train/test split; the loader uses the first 50% as train (train labels are discarded, keeping the unsupervised assumption).
  • Mostly point anomalies — poorly suited for evaluating long collective anomalies.