FAMILY_INTERNATIONAL_LEAD_LAG_DYNAMICS: Experiment Log

Overview

Testing properly aligned weekly international price lead-lag relationships between Netherlands and neighboring markets (Belgium, Germany, France). This hypothesis uses REAL DATA ONLY from BoerderijApi's international price records, with proper weekly alignment instead of flawed exact-date matching.

Hypothesis Origins

• FAMILY_CROSS_MARKET_COUPLING: CONDITIONALLY SUPPORTED (86-87%) but used contemporaneous prices
• Data Discovery: Boerderij Excel contains 438 BE, 190 DE, 152 FR weekly records now API-accessible
• Failed Exact-Date Studies: Previous cross-market analyses failed due to different reporting days
• Industry Evidence: German potato traders report 1-2 week price leadership over Dutch market
• Academic Basis: Granger (1969) causality; Engle & Granger (1987) cointegration theory

Experiment Design

• Method: Rolling-origin cross-validation
• Training Window: 365 days minimum
• Step Size: 7 days (weekly)
• Test Window: 60 days maximum
• Baselines: ALL mandatory standard baselines (persistent, seasonal_naive, ar2, historical_mean)
• REAL DATA ONLY: BoerderijApi international prices

Data Sources (REAL DATA ONLY)

• Dutch Prices: BoerderijApi - NL.157.2086 consumption potatoes - git:current
• Belgian Prices: BoerderijApi - BE.157.2086 (438 weekly records, legacy=true) - git:current
• German Prices: BoerderijApi - DE.157.2086 (190 weekly records, legacy=true) - git:current
• French Prices: BoerderijApi - FR.157.2086 (152 weekly records, legacy=true) - git:current
• Critical: Weekly alignment by week-of-year, NOT exact date matching

Experiment Runs

Variant A: German Market Leads Dutch by 1-2 Weeks

Status: Ready for implementation - Model: RandomForest, GradientBoosting, Ridge regression - Features: de_price_lag_1w, de_price_lag_2w, de_price_momentum_2w, nl_de_price_spread_1w, de_volatility_4w, cross_correlation_de_nl, lead_lag_stability_de, transport_cost_ma4, nl_price_lag_1w - Horizons: 30-day, 60-day - Mechanism: German market size advantage creates systematic 1-2 week information lead over Dutch market - Expected: >15% improvement using real DE-NL price relationships - Hypothesis: German prices lagged 1-2 weeks predict Dutch prices due to market size asymmetries and early price discovery

Variant B: French Market Leads Belgian by 2-3 Weeks

Status: Ready for implementation - Model: GradientBoosting, RandomForest, XGBoost - Features: fr_price_lag_2w, fr_price_lag_3w, be_price_lag_1w, fr_be_price_spread_2w, fr_price_momentum_3w, triangular_signal_fr_be_nl, fr_volatility_4w, be_nl_correlation_4w, transport_cost_fr_index, nl_price_lag_1w - Horizons: 30-day, 60-day - Mechanism: French market leadership transmits through Belgian processing connections to Dutch prices with compound 2-3 week lag - Expected: >12% improvement using real FR-BE-NL price chain - Hypothesis: Triangular causality chain (FR→BE→NL) outperforms direct relationships

Variant C: Multi-Market Granger Causality Network

Status: Ready for implementation - Model: Ensemble (GradientBoosting 0.4, RandomForest 0.4, VectorAutoregression 0.2) - Features: de_granger_signal, be_granger_signal, fr_granger_signal, de_optimal_lag, be_optimal_lag, fr_optimal_lag, causality_strength_de/be/fr, reverse_causality_nl_de, network_centrality_nl, dominant_lead_market, multi_market_momentum, causality_stability_4w, cointegration_residual, nl_price_lag_1w - Horizons: 30-day, 60-day - Complexity: Multi-directional causality network with optimal lag selection and cointegration - Expected: >18% improvement through systematic directional causality analysis - Hypothesis: Different optimal lag structures across market pairs with network effects providing superior forecasting

Statistical Tests

• Diebold-Mariano test with Harvey-Leybourne-Newbold correction vs strongest baseline
• TOST equivalence test with SESOI = 15% improvement (information advantage from lead-lag)
• Granger causality tests for all market pairs (DE→NL, FR→BE, BE→NL, reverse tests)
• Johansen cointegration test for long-run equilibrium relationships
• FDR correction for multiple comparisons across variants
• ALL 4 mandatory standard baselines: persistent, seasonal_naive, ar2, historical_mean (strongest baseline comparison)

Weekly Alignment Methodology (CRITICAL INNOVATION)

• ISO Week Matching: Align by week-of-year, NOT calendar dates
• Reporting Schedule Correction: BE Tuesday, DE Wednesday, FR Thursday, NL Friday
• Week-Ending Aggregation: Standardize all markets to week-ending prices
• Maximum lag: 4 weeks for causality testing
• Cross-correlation: 52-week rolling windows for relationship stability
• Optimal lag selection: AIC criterion for each market pair
• Causality significance: p < 0.05 for directional relationships

Verdicts

(Experiments not yet run)

HE Notes

• Created 2025-08-18 leveraging newly discovered international price data
• Corrects fundamental flaw in previous studies (exact-date matching)
• Weekly alignment crucial - markets report on different weekdays
• First proper lead-lag analysis with actual international data
• SESOI set at 25% due to information advantage from leading markets
• All variants use ONLY REAL DATA from BoerderijApi

Decision Log

(To be updated after experiment completion)