PREVIEW · static UI mock · screen 07 / 32 · diff + gate + Approve/Reject · all screens
High impact Awaiting review ● Gate passing

Detect seasonal anomaly in winter footwear demand 6+ weeks earlier

A cluster of 14 traces shows the agent missed the early-November demand spike for winter boots in the Pacific Northwest region — markdown alerts fired 4 weeks too late, exposing approximately $340k in markdown risk. The proposed change adds an early-signal detector that combines NOAA weather forecast deltas with prior-year YoY demand-shape comparison, firing markdown alerts as soon as both signals diverge by more than 2σ.

Source Cluster #ce-2284 · 14 traces
Skill seasonal-anomaly-detector
Proposed by Improvement agent (auto)
Created 2h ago · 2026-05-02 09:14 PT

Skill diff · seasonal-anomaly-detector v3 → v4

Current · v3 Active in prod
# seasonal-anomaly-detector v3 When forecasting weekly demand for SKU-region pairs: 1. Compute prior-12-month baseline. 2. Apply seasonal index from regional norms (NRF 2024). 3. If forecast deviates from baseline by >40%, flag for review. 4. Emit markdown_alert if forecast > inventory × 1.5. Skills retained across turns: - regional_norms - prior_baseline
Proposed · v4 +5 lines · −1 line
# seasonal-anomaly-detector v4 When forecasting weekly demand for SKU-region pairs: 1. Compute prior-12-month baseline. 1a. Compare prior-12 vs prior-24 month YoY shape — flag shape divergence. 1b. 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). Skills retained across turns: - regional_norms - prior_baseline - weather_signals

Why this change

The current logic only fires markdown_alert after forecast exceeds inventory by 50%. By that time, the demand spike has already begun in real-world signals. The current 4-week markdown lag is the root cause of the $340k exposure on this cluster.

The proposed change adds two early-signal inputs that fire before the spike materializes in demand data: weather-forecast deltas (which pre-date the spike by 2-3 weeks for cold-weather SKUs) and YoY shape divergence (which catches the structural shift earlier than absolute-value thresholds).

Source trace excerpt · 1 of 14

trace_id tr-9f4c12-2026-04-21 View all 14 traces →
> forecast_demand(sku=BOOT-W42, region=PNW, week=2026-W45) < { forecast: 312, confidence: 0.74, inventory: 1850 } > outcome (recorded 2026-W49): < { actual: 1847, markdown_alert_fired: 2026-W49 } > retrospective signal (would-have-fired with v4): < { weather_anomaly_2026-W41: cold spike +12°F below avg } < { yoy_shape_divergence_2026-W42: 3.4σ vs prior-24 baseline } < markdown_alert_v4 would have fired: 2026-W42 (7 weeks earlier)