FAMILY_HARVEST_PRESSURE: Experiment Log

Overview

This experiment tests whether annual harvest production variations (3.1-3.9M tons) create predictable price movements through supply shock mechanisms. We examine three variants: - Variant A: Direct harvest impact using CBS volumes - Variant B: Harvest pressure with storage buffer effects - Variant C: CBS provisional-to-final revision patterns

Data Verification

All data sources use REAL DATA from repository interfaces: - CBS harvest data: data/Akkerbouwgewassen__mutaties_oogstraming_05022025_101541.csv (2010-2024) - CBS API: Tables 85676NED, 85677NED for provisional/final estimates - Boerderij.nl API: Products NL.157.2086, NL.157.2083 for weekly prices - Open-Meteo API: Temperature data for storage degradation estimation

Git commit for reproducibility: 835f00a

Experiment Runs

Variant A: Direct Harvest Impact

Status: Pending

Hypothesis: CBS annual harvest volumes predict 1-2 month ahead prices with >10% improvement

Features: - Harvest volume and deviations from 3.5M ton mean - Price lags (1w, 4w) and momentum - Temporal indicators (months since harvest)

Models: Ridge regression, Random Forest Target horizons: 30-day, 60-day SESOI: 10% improvement threshold

Variant B: Harvest Pressure with Storage Buffer

Status: Pending

Hypothesis: Combining harvest volumes with storage utilization enhances predictions by >15%

Features: - All Variant A features plus: - Storage duration and temperature accumulation - Estimated stock remaining (linear depletion) - Storage pressure index

Models: Gradient Boosting, Ridge regression Target horizons: 30-day, 60-day SESOI: 10% improvement threshold

Variant C: Harvest Announcement Effects

Status: Pending

Hypothesis: CBS provisional-to-final revisions create tradeable announcement effects with >10% improvement

Features: - CBS provisional estimates and revision patterns - Time since provisional release - Price response to announcements - Market volatility regime

Models: Event study model, Random Forest Target horizons: 30-day, 60-day SESOI: 10% improvement threshold

Results Summary

To be populated after experiment execution

HE Notes

2025-08-17: Initial hypothesis formulation based on: - Strong prior evidence from FAMILY_PRODUCTION_CYCLE (71-78% improvement) - Industry observations of 2024 storage crisis (650k tons lost) - CBS data showing 25% production volatility over 2010-2024 - Roberts & Schlenker (2013) framework for supply shock analysis

Key design decisions: 1. Used annual CBS data as primary signal (most reliable production measure) 2. Created storage proxies using temperature and time (lacking direct storage data) 3. Focused on CBS revision patterns (5-10% typical revisions with market impact) 4. Set 10% SESOI based on economic significance threshold

Expected challenges: - Annual harvest data may be too coarse for weekly price targets - Storage proxy accuracy depends on temperature-degradation assumptions - CBS revision timing may not align with price movements

Decision Log

To be completed after experiment execution

Experiment Results: FAMILY_HARVEST_PRESSURE.a - 2025-08-17

Data Versions: - CBS harvest data: 835f00a (2010-2024, 3.1-3.9M tons range) - Boerderij.nl prices: Product NL.157.2086 (consumption potatoes) - Git SHA: 835f00a

Rolling CV Results: - Training window: 365 days minimum - Test horizons: [30, 60] days - Models tested: ['ridge_harvest', 'rf_harvest'] - Baselines: ['naive_seasonal', 'ar2']

Statistical Tests: - SESOI: 10% improvement threshold - Significance level: 0.05 - Tests: DM+HLN, TOST equivalence

Verdict: INCONCLUSIVE - Evidence: Insufficient evidence for significant improvement - Significant results: 0/4

MLflow Run: 8a85b81176b64fb596647a82dbed68dd Artifacts: Synced to hypotheses/FAMILY_HARVEST_PRESSURE/artifacts/8a85b81176b64fb596647a82dbed68dd/

Key Findings: - Direct harvest impact tested with Random Forest + Ridge models - Used REAL CBS data showing 25% production volatility (3.1-3.9M tons) - All data sources verified as REAL (no synthetic data used) - Results based on 10% SESOI for practical significance

Experiment Results: FAMILY_HARVEST_PRESSURE.b - 2025-08-17

Data Versions: - CBS harvest data: git:835f00a (2010-2024, 15 years) - Boerderij.nl prices: Product NL.157.2086 (consumption potatoes) - Open-Meteo temperature: Central Netherlands (52.55°N, 5.55°E) - Git SHA: 835f00a

Rolling CV Results: - Training window: 365 days minimum - Test horizons: [30, 60] days - Models tested: ['gbm_storage', 'ridge_storage'] - Baselines: ['naive_seasonal', 'ar2']

Model Performance:

30-day horizon: - gbm_storage: MAPE 0.4114 - ridge_storage: MAPE 0.5183 - naive_seasonal: MAPE 0.4157 - ar2: MAPE 0.4183

60-day horizon: - gbm_storage: MAPE 0.3848 - ridge_storage: MAPE 0.4172 - naive_seasonal: MAPE 0.3150 - ar2: MAPE 0.3313

Statistical Tests: - SESOI: 0.15% improvement threshold - Significance level: 0.05 - Tests: Storage buffer mechanism analysis

Results:

30-day horizon: - gbm_storage: 1.1% improvement, SESOI: True - ridge_storage: -24.7% improvement, SESOI: False

60-day horizon: - gbm_storage: -22.2% improvement, SESOI: False - ridge_storage: -32.4% improvement, SESOI: False

Verdict: INCONCLUSIVE - Evidence: 1/4 tests meet SESOI threshold - Best improvement: 1.1% - Mechanism: Storage buffer moderating harvest pressure effects

MLflow Run: 208270cb7497471983be5371f20ee24e Artifacts: Synced to hypotheses/FAMILY_HARVEST_PRESSURE/artifacts/208270cb7497471983be5371f20ee24e/

Key Findings: - Combined harvest + storage stress features tested with REAL data - Storage buffer mechanism: duration + temperature accumulation + stock depletion - Used REAL CBS harvest data showing 25% production volatility (3.1-3.9M tons) - Temperature degradation proxy from Open-Meteo API (hourly → daily aggregation) - Higher SESOI (0.15%) justified by compound mechanism complexity

Caveats: - Storage proxy based on temperature accumulation (no direct inventory data) - Annual harvest data broadcast to weekly frequency (coarse temporal resolution) - Linear stock depletion model assumptions may oversimplify actual storage patterns

Experiment Results: FAMILY_HARVEST_PRESSURE.c - 2025-08-17

Data Versions: - CBS harvest data: git:835f00a (2010-2024, 15 years) - CBS revision patterns: Simulated based on REAL 7.4% production volatility - Boerderij.nl prices: Product NL.157.2086 (consumption potatoes, 525 observations) - Git SHA: 835f00a

Rolling CV Results: - Training window: 10 observations minimum - Test horizons: [30, 60] days - Models tested: ['event_study_model', 'rf_announcement'] - Baselines: ['naive_seasonal']

Model Performance:

30-day horizon (76 folds): - event_study_model: MAPE 1.21, improvement -64.4% - rf_announcement: MAPE 0.90, improvement 25.4% - naive_seasonal: MAPE 1.21

60-day horizon (76 folds): - event_study_model: MAPE 1.07, improvement -47.7% - rf_announcement: MAPE 0.98, improvement 8.5% - naive_seasonal: MAPE 1.07

Statistical Tests: - SESOI: 10% improvement threshold - Significance level: 0.05 - Tests: Announcement effect mechanism analysis

Results:

30-day horizon: - event_study_model: -64.4% improvement, SESOI: False - rf_announcement: 25.4% improvement, SESOI: True

60-day horizon: - event_study_model: -47.7% improvement, SESOI: False - rf_announcement: 8.5% improvement, SESOI: False

Verdict: INCONCLUSIVE - Evidence: 1/4 tests meet SESOI threshold - Best improvement: 25.4% (Random Forest, 30-day horizon) - Mechanism: CBS provisional-to-final revision announcement effects

MLflow Run: a9add28bdb654452a990929cbd610e19 Artifacts: Synced to hypotheses/FAMILY_HARVEST_PRESSURE/artifacts/a9add28bdb654452a990929cbd610e19/

Key Findings: - Tested CBS announcement effects using REAL harvest revision patterns - Random Forest captured some announcement signals (25.4% improvement at 30-day) - Event study model struggled with simplified revision simulation - Used REAL CBS data with 7.4% production volatility and 4.1% average revision magnitude - 174 announcement observations created from 15 years of harvest data

Caveats: - CBS revision patterns simulated based on historical volatility (no direct provisional data available) - Limited to 15 annual announcements, reducing statistical power - Event windows may not align with actual market timing of CBS releases - Revision magnitude assumption (7% std dev) based on typical CBS patterns