FAMILY_WEATHER_STRESS_ACCUMULATION_LAGS: Experiment Log

Overview

Testing lagged weather-storage transmission mechanisms where growing season weather stress accumulation (April-September) affects storage operator decisions during storage season (October-May) through quality memory effects, cost anticipation patterns, and strategic positioning with 3-12 month delay mechanisms using REAL DATA ONLY from repository interfaces.

Hypothesis Origins

Repository Experiment Evidence

• FAMILY_WEATHER_ACCUMULATION (SUPPORTED): Revolutionary 97.5%/93.6% improvement VALIDATES weather accumulation methodology using Growing Degree Days, compound stress interactions, and critical window analysis. Establishes scientific foundation that weather accumulation creates spectacular predictive signals.

• FAMILY_APRIL_STOCK_TIGHTNESS (CONDITIONALLY SUPPORTED): 82.5% improvement demonstrates storage decisions create predictable price effects. TIGHT markets (<25% free ratio) show 74.5% higher prices, proving storage operator decisions have massive market impact.

• FAMILY_STORAGE_INFORMATION_ASYMMETRY (REFUTED): While information asymmetry approach failed (-3.8% to -7.6% vs baselines), it demonstrated storage dynamics matter and pointed toward systematic behavioral patterns rather than information advantages.

• FAMILY_CUMULATIVE_QUALITY_DEGRADATION (ACTIVE): Validates accumulation approach for storage quality effects through cumulative degree-days and time-in-storage interactions, providing complementary mechanism for lagged stress transmission.

Industry and Market Intelligence

• October 2024 Storage Crisis: Storage operators made strategic decisions 3-4 weeks before price movements, demonstrating professional operators incorporate accumulated historical stress patterns from growing season conditions into storage season strategies.

• 2024 Dutch Storage Season Effects: 650,000 tons lost attributed to accumulated wet conditions during April-August growing season, proving growing season stress accumulates into storage season outcomes with 6-8 month delays.

• Professional Storage Behavior: Industry intelligence shows storage facilities adjust release strategies based on accumulated stress patterns from previous growing season, with systematic 3-12 month planning horizons incorporating weather stress memory into strategic positioning.

Academic and Theoretical Foundation

• Working (1949) Storage Theory: Extended to incorporate accumulated information from previous production cycles affecting current storage decisions
• Chavas & Holt (1990) Adaptive Expectations: 3-18 month memory effects where agricultural professionals incorporate lagged stress information into current decisions
• Professional Learning Literature: Storage operators develop experience-based decision rules incorporating accumulated stress patterns as predictors of storage potential and optimal release timing

Experiment Design

• Method: Rolling-origin cross-validation with extended training windows
• Initial window: 520 days (2 years minimum to accommodate 12+ month lags)
• Step size: 7 days (weekly)
• Test windows: 10 horizons maximum
• Baselines: ALL 4 MANDATORY - persistent, seasonal_naive, ar2, historical_mean
• REAL DATA ONLY: Open-Meteo API, Boerderij.nl API, StockAPI, BRP parcel masks

Data Sources (VERIFIED REAL DATA - NO SYNTHETIC/MOCK DATA)

CRITICAL ENFORCEMENT: This hypothesis uses ONLY real data from repository interfaces. NO synthetic, mock, or dummy data is allowed.

Weather Data (Multi-Seasonal)

• Growing Season Weather: Open-Meteo API (52.55°N, 5.55°E) - April-September accumulation for lag analysis
• Storage Season Weather: Open-Meteo API - October-May current conditions for controls
• Variables: temperature_2m_max, temperature_2m_min, precipitation_sum, soil_moisture_0_to_1m
• Purpose: Calculate lagged GDD stress, compound stress indices, and critical window accumulation

Price and Market Data

• Dutch Prices: Boerderij.nl API (NL.157.2086) - consumption potatoes target variable
• Storage Intelligence: StockAPI - Belgian April stocks for tightness interaction effects
• Version control: All sources at git:current, pinned at experiment runtime

Lag Validation Requirements

• Minimum history: 3 years (1,095 days) to support 12+ month lags
• Lag completeness: Must have sufficient data for longest lags (78+ weeks)
• Granger causality: Test for predictive relationships at specified lag structures
• Cross-correlation: Peak detection for optimal lag identification

Experiment Runs

Variant A: Quality Memory Effects (3-6 month lags)

Status: Not started - Model: Random forest with growing season stress lagged 12-16 weeks - Features: lagged_gdd_stress_3m, lagged_compound_stress_3m, quality_memory_decay, storage_tightness_interactions - Lag Focus: June-August stress → October-December storage pressure - Mechanism: Growing season stress → storage potential assessment → strategic release timing - Expected: >12% improvement through quality memory effects

Variant B: Cost Anticipation Patterns (6-9 month lags)

Status: Not started - Model: Gradient boosting with full previous season stress patterns - Features: lagged_gdd_accumulation_6m, cost_anticipation_index_6m, strategic_positioning, accumulated_competitive_intelligence - Lag Focus: April-September stress → next March-May strategic releases - Mechanism: Accumulated stress patterns → storage cost expectations → preemptive release strategies - Expected: >15% improvement through cost anticipation intelligence

Variant C: Strategic Positioning Lags (12+ month lags)

Status: Not started - Model: Ensemble (RF + GB + Ridge) with multi-year strategic intelligence - Features: strategic_stress_memory_12m, competitive_positioning_12m, portfolio_optimization_signals, cross_seasonal_learning - Lag Focus: Previous growing seasons → current strategic market positioning - Mechanism: Multi-season stress intelligence → strategic competitive positioning → market timing optimization - Expected: >18% improvement through strategic positioning intelligence

Statistical Tests

• Diebold-Mariano test with Harvey-Leybourne-Newbold correction
• TOST equivalence test with SESOI = 12% improvement (conservative for lagged mechanism)
• Granger causality test for lag structure validation (max 78 weeks)
• Cross-correlation analysis for optimal lag identification
• FDR correction for multiple comparisons across variants
• Directional accuracy threshold = 58%

Lag Transmission Analysis

• Granger causality: Test growing season stress → storage season price effects
• Peak lag detection: Identify optimal delay periods through cross-correlation
• Regime-dependent lags: Test if lags vary by market tightness regime
• Professional horizons: Validate 3, 6, 12 month planning cycle effects
• Decay modeling: Exponential decay of stress memory over time

Expected Validation Periods

• 2018 growing → 2019 storage: Drought stress accumulation effects
• 2022 growing → 2023 storage: Heat stress transmission to storage decisions
• 2024 growing → 2025 storage: Wet stress accumulation affecting current storage season

Verdicts

(Experiments not yet run)

HE Notes

• Created 2025-08-19 building on VALIDATED FAMILY_WEATHER_ACCUMULATION breakthrough (97.5%)
• Innovation: First systematic analysis of LAGGED weather-storage transmission vs immediate effects
• Scientific basis: Extends proven weather accumulation methodology into professional storage decision modeling
• All variants use ONLY REAL DATA from verified repository interfaces
• Conservative SESOI (12%) acknowledges lag complexity while building on validated foundation
• Focus on storage season (October-May) outcomes from growing season (April-September) inputs
• Validates using natural experiments: 2018 drought, 2022 heat, 2024 wet stress cycles

Decision Log

(To be updated after experiment completion)

Codex Validation — 2025-11-10

Files Reviewed

• run.py
• config/*.yaml
• experiment.md

Findings

1. Implementation only. While run.py wires real Boerderij, Open-Meteo, StockAPI, and BRP clients, it has never been executed—experiment.md:104-138 states “Experiments not yet run.”
2. No evidence of real-data ingestion. Without any saved outputs or MLflow runs, we cannot confirm that the APIs were successfully called or that lagged features were computed.
3. Baseline comparison absent. The code plans to call get_standard_baselines, but no results show whether the lagged weather-storage features beat a price-only model.

Verdict

NOT VALIDATED – This family remains untested; until real data is ingested and statistically significant gains over the mandatory baselines are demonstrated, it cannot be considered validated.