PREVIEW · static UI mock · screen 18 / 32 · skill detail · content + retention contract · all screens
seasonal-anomaly-detector
v4 ● Active in prod forecasting anomaly-detection weather
open agent UI →
WorkflowDemand prediction
OwnerMaria Chen
Created2025-09-14
Last update2h ago · v3 → v4
Used by1 workflow · forecast loop
Eval coverage12 cases
Content Versions 4 Used by 1 Eval cases 12 Audit

Skill content · v4

seasonal-anomaly-detector v4SKILL.md · 28 lines · 1.4 KB

Purpose

Forecast weekly demand at SKU-region level and emit early markdown alerts when seasonal demand diverges from baseline.

Procedure

When forecasting weekly demand for SKU-region pairs:

  1. Compute prior-12-month baseline.
    1. Compare prior-12 vs prior-24 month YoY shape — flag shape divergence.
    2. Cross-reference NOAA regional weather forecast deltas.
  2. Apply seasonal index from regional norms (NRF 2024).
  3. If forecast deviates from baseline by >40%, flag for review.
  4. Emit markdown_alert as soon as YoY shape diverges by >2σ AND weather signal supports (precipitation/temp anomaly).

Inputs (tools the agent must call)

  • query_sap_sales(sku, region, years=4) — historical sales
  • fetch_noaa_weather(region, weeks_ahead=12) — weather forecast deltas
  • get_inventory_status(sku, store) — current inventory

Outputs

  • fire_markdown_alert(sku, region, severity, recommended_discount_bp)
  • Forecast value with confidence interval and citations

— end of skill · retention contract follows below —

Retention contract

What state this skill keeps across turns. Violations are caught by eval case #126 — the agent must re-fetch weather_signals each turn or the test fails.
regional_norms Keep across turns within session · invalidate on session end · NRF data is stable per quarter eval #98
prior_baseline Keep until SKU context changes · re-compute when forecasting a new SKU-region eval #97
weather_signals Re-fetch each turn · do not retain stale · weather updates hourly eval #126

Recent eval results

Last full run: 12 / 12 passing · 2h ago · 18s · All-time best: 12 / 12 (v4) · Confidence: high · 0 retention violations