Signals, segments & time series

Signals

A signal is a named time series: a key plus a stream of timestamped values.

daily_sales      → (2026-07-13, 12100.0), (2026-07-14, 12841.5), …
web_traffic      → (2026-07-13, 48211),   (2026-07-14, 51037), …
machine_temp_c   → (2026-07-14T09:00, 61.2), (2026-07-14T09:05, 61.9), …

Signals are created implicitly: push a value to a new key and the signal exists. There's no schema to declare. Data arrives three ways:

• Push — POST /v1/signal/push, one value at a time (great for live data).
• Ingestion — connectors, file uploads, and scheduled jobs that map columns to signal keys (Ingestion).
• Webhooks — external providers push to a per-source HMAC-verified URL.

Signals carry lightweight metadata you can manage through the API: display names, units, and group assignments — all covered in Signals.

Segments

A segment is a curated view over signals: which keys, over what window, at what resolution, with what normalization. It's the exact table a model sees at training time.

Segments matter because raw signals are rarely model-ready. A segment pins down the decisions — alignment, interpolation, scaling — so that training and inference see the same data shape.

You'll meet segments in three places:

• Pipelines bind a segment to a model — "train this model on this data."
• Goals build and maintain their own segment automatically.
• Segments is where you create and manage them directly — including normalization strategies and engineered features.

Analysis

Signals support built-in analysis before you ever train anything: correlation between signals, relevance ranking against a target, causality tests, and workspace-wide analysis jobs. Use these to understand your data — or skip straight to a goal, which runs discovery for you.