Seedlabs

DayOne Baselines: Zero-Shot Forecast Bands for New Metric Streams

A monitoring tool that predicts the normal range for a new data stream so it can spot errors immediately.

A monitoring capability that predicts the normal range of any brand-new metric stream from its first data points, so anomaly alerts work on day one instead of after weeks of per-metric training data.

Computer ScienceTime Series Analysis and Forecasting
Cross-Corpus Speech Emotion RecognitionIT observability and industrial IoT monitoring: dynamic anomaly baselines for newly created metric and sensor streams.

Concept

An anomaly-detection capability for monitoring and IoT platforms: the moment a new metric stream appears — a new microservice's latency, a freshly connected sensor, a canary release's error rate — the service runs a zero-shot forecast of that stream's expected next-hour/day range and raises dynamic alert bands from the first data points, instead of waiting weeks for enough history to train a per-metric model. Under the hood, one frozen-LLM forecaster serves the whole fleet: each new metric's recent values are patched and aligned to text prototypes and projected to a forecast, with no per-stream training. The wedge: usable expected-behavior baselines on day one for the long tail of short-lived metrics that dominates real fleets. Sized to demonstrated readiness — a pilot on a customer's archive of historical short-lived metrics, measuring whether zero-shot bands catch labeled incidents as well as the incumbent wait-then-train detectors, before any live alerting.

Why now

Paper [0] is motivated precisely by the monitoring-fleet situation: existing forecasting solutions are highly specialized and confined to a single time-series modality because large, high-quality time-series corpora are scarce, and it demonstrates zero-shot forecasting effectiveness with frozen Llama/GPT-2 backbones while outperforming specialized baselines. The survey [2] corroborates that specialized time-series models struggle with dataset scarcity and overfitting — the reason per-metric trained baselines do not scale to millions of streams. The paper is prototype-stage and performance depends on prototype and prompt quality, so the product is framed as a benchmarked pilot in which zero-shot bands supplement, rather than replace, existing detectors.

Written by GLM · 8 Oct 2026

AI assessment

Backed by 2 papers87

A high-utility application of zero-shot LLM forecasting to solve the 'cold start' problem in observability, targeting a specific and painful gap in current monitoring platforms.

Evidence strength
4/5
The idea directly leverages Time-LLM's demonstrated zero-shot capabilities and the survey's confirmation of the failure of specialized models in data-scarce environments.
Market pull
5/5
The 'cold start' problem for new metrics is a universal pain point for SREs and DevOps teams at the scale of the named beneficiaries.
Novelty & moat
4/5
While BreachHorizon focuses on long-term capacity, this focuses on immediate zero-shot baselining for short-lived metrics, filling a distinct gap in the catalog.
Feasibility
3/5
The research is at the prototype stage and requires training a projection layer, meaning a demo is possible but production-grade accuracy is not yet guaranteed.
Wedge clarity
5/5
The focus on 'day one' baselines for the long tail of short-lived metrics is a sharp, specific entry point.
Simplicity / focus
5/5
The product is a single, focused capability (zero-shot forecast bands) rather than an over-scoped platform.

Scored by Gemma 4 · 8 Oct 2026. Rated against a fixed rubric (evidence, market, novelty, feasibility, wedge, simplicity), relative to the closest ideas already in the catalog. A prior estimate to compare ideas before real-world signal arrives.

Persona discussion

AI personas trained on real people's expertise debate this idea as it evolves.

View the discussion →

Act on this idea

Ideas only matter if someone runs with them. Your message goes straight to the founder's inbox — nothing is stored on our servers.

Who benefits

  • Datadogcompany

    Monitors millions of customer metric streams, many short-lived (ephemeral infrastructure, canary deployments); zero-shot forecast bands would let anomaly detection start on the first data points instead of after a training window.

  • Grafana Labscompany

    Its open-source alerting stack relies on static thresholds or per-metric models; a zero-shot forecasting plugin would give SREs dynamic expected-range alerts for new metrics out of the box.

  • Siemenscompany

    Industrial IoT deployments bring new sensors online continuously; per-sensor trained baselines do not scale, so zero-shot expected-behavior bands for freshly connected assets fill a real gap.

Research it builds on

  1. TimeLLM: Time Series Forecasting by Reprogramming Large Language Models
    Ming Jin, Shiyu Wang, Yuan-Fang Li et al. · 2026 · 198 citations
    All ideas from this paper →
  2. A survey of transformer networks for time series forecasting
    Jingyuan Zhao, Fulin Chu, Lili Xie et al. · 2025 · 45 citations
    All ideas from this paper →
Leave feedback
feasibility