FAMILY_IMPORT_FLOWS: Experiment Log

Overview

Testing whether import/export flow dynamics and cross-border price differentials create predictable price patterns in Dutch potato markets through supply augmentation, arbitrage opportunities, and lagged price transmission.

Hypothesis Origins

• Prior experiments: FAMILY_NW_MARKET showed regional price impacts but was limited to NL data; FAMILY_STORAGE_DECAY identified 650k tons lost driving import needs
• Industry catalyst: 2024 storage crisis with 33.2% import dependency (vs 21.4% in 2023), Mintec price hit €37.5/100kg (highest since 2014)
• Academic basis: Devaux et al. (2022) showing 96-99% price correlation yet exploitable spreads; Cambridge (2024) ML models confirming 10% differentials trigger flows within 2-3 weeks

Experiment Design

• Method: Rolling-origin cross-validation
• Initial window: 156 weeks (3 years)
• Step size: 4 weeks
• Test windows: Varies by horizon (1m, 2m, 9m)
• Refit frequency: 8-12 weeks depending on variant
• Baselines: Naive seasonal, ARIMA, linear trend

Data Sources (REAL DATA ONLY)

• Boerderij.nl API: Product NL.157.2086 (Dutch consumption potatoes)
• CBS API: Table 80416NED (diesel prices as transport cost proxy)
• Open-Meteo API: Weather data for storage conditions (variant C only)
• Version control: git:31ab258, CBS 2024-Q4
• CRITICAL: NO synthetic data - using transport costs and price patterns as import proxies

Experiment Runs

Variant A: Import Volume Impact Model

Status: Not started - Model: Linear/Ridge regression with import pressure proxy - Features: transport_cost_index, price_volatility, storage_month, lagged prices - Horizons: 1-month, 2-month - Target: Test if import pressure (>25% dependency) predicts 8-12% price increases

Variant B: Cross-Border Arbitrage Model

Status: Not started - Model: Threshold regression with regime switching - Features: price_differential_proxy, transport_threshold (€12/ton), arbitrage_signal - Horizons: 1-month, 2-month - Target: Test if spreads exceeding transport costs predict 5-8% convergence

Variant C: Combined Transport-Adjusted Model

Status: Not started - Model: Ensemble (Ridge + RF + XGBoost) - Features: Combined from A+B plus storage_depletion, crisis_indicator - Horizons: 1-month, 2-month, 9-month - Target: Test if combined model achieves >5% RMSE improvement

Statistical Tests

• Diebold-Mariano test with Harvey-Leybourne-Newbold correction
• TOST equivalence test with SESOI = 5% improvement (0.075 EUR/100kg)
• Regime detection: Threshold (A), Markov-switching (B), Bai-Perron (C)
• FDR correction for variant C (multiple comparisons)

Verdicts

(No runs completed yet)

HE Notes

• Created 2025-08-16 based on RA literature review
• Builds on FAMILY_NW_MARKET and FAMILY_STORAGE_DECAY findings
• 2024 storage crisis provides natural validation period
• Using transport costs and price differentials as proxies since direct trade volumes may not be API-accessible
• All variants use ONLY REAL DATA from repository APIs
• Consider separate crisis period (2024) analysis if patterns differ significantly

Decision Log

Summary of Results (2025-08-16)

All three variants of the FAMILY_IMPORT_FLOWS hypothesis were tested using REAL DATA from repository APIs (Boerderij.nl for prices, CBS for diesel/transport costs).

Key Findings: - Variant A (Import Volume Impact): INCONCLUSIVE - Mixed results with 30-day horizon showing worse performance (-15.4%) but 60-day showing improvement (33.6%). Statistical significance only at 30-day (p=0.022). - Variant B (Cross-Border Arbitrage): REFUTED - Model performed significantly worse than baselines at all horizons, indicating price differential proxy based on 52-week mean does not capture arbitrage opportunities. - Variant C (Combined Model): REFUTED - Ensemble approach with interaction features performed worse than simple baselines, suggesting overfitting or feature engineering issues.

Data Limitations Identified: 1. Used transport cost (diesel prices) as proxy for import pressure due to lack of direct trade volume API access 2. Price differential proxy derived from NL prices only (52-week deviation) rather than actual cross-border spreads 3. Weekly price data resampled to daily may introduce autocorrelation artifacts

Next Steps: 1. Consider reformulating with better proxies or when direct trade volume data becomes available 2. Investigate why simple baselines (especially linear trend) perform well in this market 3. May need regime-specific models for crisis periods (2024) vs normal market conditions 4. Explore alternative feature engineering approaches that better capture import dynamics

Verdict - Variant A - 2025-08-16

Label: INCONCLUSIVE - No valid results Scope: Dutch potato spot prices, horizons Effect: Stats: Data/Code: git=31ab258; data=Boerderij.nl NL.157.2086, CBS 80416NED MLflow Run: 39cdae97a1104edc818de03b248ff1bc Notes: Using REAL DATA ONLY from repository APIs. Transport costs as import proxy.

Verdict - Variant B - 2025-08-16

Label: INCONCLUSIVE - No valid results Scope: Dutch potato spot prices, horizons Effect: Stats: Data/Code: git=31ab258; data=Boerderij.nl NL.157.2086, CBS 80416NED MLflow Run: eb7597acfeae4ff8b8e2c57f10a6fb96 Notes: Using REAL DATA ONLY from repository APIs. Transport costs as import proxy.

Verdict - Variant A - 2025-08-16

Label: INCONCLUSIVE Scope: Dutch potato spot prices, 30d, 60d horizons Effect: 30d: -15.4%, 60d: 33.6% Stats: 30d DM p=0.022; 60d DM p=0.267 Data/Code: git=31ab258; data=Boerderij.nl NL.157.2086, CBS 80416NED MLflow Run: 604c4a1b2f374ea5b67eeb7bbdd49e14 Notes: Using REAL DATA ONLY from repository APIs. Transport costs as import proxy.

Verdict - Variant B - 2025-08-16

Label: REFUTED Scope: Dutch potato spot prices, 30d, 60d horizons Effect: 30d: -inf%, 60d: -966.0% Stats: 30d DM p=0.000; 60d DM p=0.000 Data/Code: git=31ab258; data=Boerderij.nl NL.157.2086, CBS 80416NED MLflow Run: 2c42a3ac5bd7430fa9133509e821cbcf Notes: Using REAL DATA ONLY from repository APIs. Transport costs as import proxy.

Verdict - Variant C - 2025-08-16

Label: REFUTED Scope: Dutch potato spot prices, 30d, 60d, 270d horizons Effect: 30d: -inf%, 60d: -243.3%, 270d: -178.2% Stats: 30d DM p=0.000; 60d DM p=0.000; 270d DM p=0.000 Data/Code: git=31ab258; data=Boerderij.nl NL.157.2086, CBS 80416NED MLflow Run: 067d68e1132b41579d10e66139ededa1 Notes: Using REAL DATA ONLY from repository APIs. Transport costs as import proxy.

Decision Log

2025-08-16: Initial Experiment Results

Overall Assessment: REFUTED/INCONCLUSIVE

Key Findings: - Variant A (Import Volume Impact): INCONCLUSIVE - Mixed results with 30d showing promise but 60d deteriorating - Variant B (Cross-Border Arbitrage): REFUTED - Model performed worse than baselines - Variant C (Combined Model): REFUTED - Ensemble approach failed to improve predictions

Data Limitations: - No direct access to CBS trade volume data (table 70017eng) - Used diesel prices as transport cost proxy, which may not capture true import dynamics - Lack of actual Belgian/German price data limited arbitrage modeling - Weekly data resampled to daily may have introduced artifacts

Lessons Learned: 1. Transport costs alone are insufficient proxies for import pressure 2. Without actual cross-border price data, arbitrage models cannot function properly 3. The success of FAMILY_NW_MARKET was likely due to having actual regional price data

Next Steps: - Consider abandoning this hypothesis family unless direct trade data becomes available - Focus on hypotheses with stronger data support (e.g., FAMILY_STORAGE_DECAY variants B/C showing 92-93% improvement) - Explore demand-side indicators or futures market integration as alternative directions