FAMILY_TRANSPORT_NETWORK_OPTIMIZATION: Experiment Log

Overview

Testing transport network optimization algorithms for Dutch potato price forecasting through real-time diesel arbitrage networks, multi-modal transport optimization, and dynamic routing network effects. This hypothesis builds on FAMILY_DIESEL_CORRELATION (REFUTED with 48-84% worse performance) by advancing beyond simple correlation to sophisticated transport network optimization intelligence.

Hypothesis Origins

• FAMILY_DIESEL_CORRELATION: REFUTED with 48-84% worse performance than baselines, but established diesel price importance - this family advances to transport network optimization
• FAMILY_CROSS_MARKET_COUPLING: CONDITIONALLY SUPPORTED with 86.8% improvement using cross-border transmission, validating international arbitrage mechanisms
• FAMILY_EUROSTAT_TRANSPORT_ARBITRAGE: Active testing of €12/ton threshold validation using Eurostat transport cost indices
• Industry catalyst: 2024 storage crisis (33.2% import dependency) revealed transport network critical importance
• Academic basis: Transport network optimization theory, multi-modal economics, network effects in agricultural markets

Experiment Design

• Method: Rolling-origin cross-validation
• Initial window: 365 days (1 year minimum)
• Step size: 7 days (weekly)
• Test windows: 60 days maximum
• Baselines: MANDATORY STANDARD BASELINES - persistent, seasonal_naive, ar2, historical_mean (from get_standard_baselines())
• REAL DATA ONLY: Eurostat API transport indices, Boerderij.nl API international prices

Data Sources (REAL DATA ONLY)

• Transport costs: Eurostat API (STS_SETU_M) - road/rail/waterway freight costs - git:current
• Diesel prices: Eurostat API (NRG_PC_204) - automotive diesel including taxes - git:current
• International prices: Boerderij.nl API (NL/BE/DE/FR.157.2086/2083) - cross-border potato prices - git:current
• Prices: Boerderij.nl API (NL.157.2086) - Dutch consumption potatoes - git:current

Experiment Runs

Variant A: Real-time Diesel Arbitrage Networks

Status: Not started - Model: Random forest with diesel-driven arbitrage features - Features: diesel_price_lag_1w, diesel_price_lag_2w, diesel_momentum_4w, transport_cost_index, nl_be_price_spread, nl_de_price_spread, nl_fr_price_spread, arbitrage_threshold_signals, diesel_volatility_4w, price_lag_1w - Horizons: 1-month, 2-month - Mechanism: Real-time diesel cost transmission through €12/ton arbitrage thresholds drives cross-border potato flows - Expected: >12% improvement over ALL 4 standard baselines

Variant B: Multi-Modal Transport Optimization

Status: Not started - Model: Gradient boosting with multi-modal optimization features - Features: road_freight_index, rail_freight_index, waterway_freight_index, modal_cost_ratios, optimal_modal_choice, modal_switching_signals, capacity_utilization_proxy, seasonal_modal_efficiency, cross_border_modal_mix, price_lag_1w - Horizons: 1-month, 2-month - Mechanism: Multi-modal transport optimization creates efficiency arbitrage affecting cross-border potato procurement patterns - Expected: >15% improvement with modal optimization intelligence

Variant C: Dynamic Routing Network Effects

Status: Not started - Model: Ensemble (RF 0.4, GB 0.4, Ridge 0.2) with network optimization algorithms - Features: network_efficiency_index, route_optimization_savings, capacity_constraint_signals, hub_centrality_effects, triangular_arbitrage_ops, network_congestion_proxy, dynamic_route_flexibility, distance_weighted_spreads, network_resilience_index, price_lag_1w - Horizons: 1-month, 2-month - Mechanism: Dynamic transport network optimization creates system-wide efficiency gains affecting potato price transmission through network effects - Expected: >18% improvement through sophisticated network algorithms

Statistical Tests

• Diebold-Mariano test with Harvey-Leybourne-Newbold correction vs ALL 4 standard baselines
• TOST equivalence test with SESOI = 12% improvement (transport networks show strong effects)
• Bai-Perron structural break test for transport optimization vs normal periods
• FDR correction for multiple comparisons across variants
• Directional accuracy threshold = 60%

Regime Analysis

• Transport optimization periods: Defined as >1 std deviation from normal transport cost patterns
• Separate performance evaluation for network stress vs normal transport regimes
• Must validate €12/ton arbitrage threshold from industry reports
• Focus on periods with transport cost volatility and cross-border arbitrage opportunities

Expected Network Effects

• Real-time diesel arbitrage: Diesel cost transmission through validated €12/ton thresholds
• Multi-modal optimization: Road/rail/waterway cost minimization creates efficiency arbitrage
• Dynamic routing: Network-wide optimization algorithms exploit system-wide efficiency gains
• Hub centrality: Netherlands hub position creates network centrality advantages

Verdicts

(Experiments not yet run)

HE Notes

• Created 2025-08-19 advancing beyond FAMILY_DIESEL_CORRELATION (REFUTED) failure
• Superior to simple diesel correlation: uses sophisticated transport network optimization vs basic price correlation
• SESOI set to 12% due to strong transport network optimization literature evidence
• All variants use ONLY REAL DATA from verified Eurostat and Boerderij repository interfaces
• Focus on 30/60-day horizons where transport optimization signals strongest
• Validates using 2022 (energy crisis), 2024 (storage crisis) natural experiments

Decision Log

(To be updated after experiment completion)

Codex Validation — 2025-11-10

Files Reviewed

• hypotheses/FAMILY_TRANSPORT_NETWORK_OPTIMIZATION/run_experiment.py
• hypotheses/FAMILY_TRANSPORT_NETWORK_OPTIMIZATION/experiment.md
• hypotheses/FAMILY_TRANSPORT_NETWORK_OPTIMIZATION/hypothesis.md

Findings

1. No mocking/patching detected – the implementation imports the real Boerderij API (run_experiment.py:18-64) and never substitutes fake data or monkey-patches runtime behavior.
2. Data provenance mismatch – although the docstrings repeatedly claim Eurostat transport feeds, every engineered feature is derived from the Dutch and international price series themselves. For example, diesel/transport proxies are just rolling volatility or spreads of nl_price (run_experiment.py:151-195). No Eurostat endpoint is queried anywhere, so the “transport network” signals are only deterministic functions of the target variable.
3. Experiment never executed – experiment.md marks every variant as “Status: Not started,” meaning there is no MLflow record, no metrics, and no statistical tests to confirm superiority over price-only baselines.
4. Price-only comparison unresolved – while run_experiment.py:218-413 sets up rolling CV with standard baselines, there is zero evidence that any variant actually ran or beat the baselines. With no persisted results, we cannot confirm that the engineered spreads outperform a plain price model.

Verdict

NOT VALIDATED – The code reads clean and uses no mocks, but no experiment was executed and all “transport” features are simply algebraic transforms of the same price series being forecast. There is no proof that the proposed variables add signal beyond price-only baselines, so this family remains unvalidated.