FAMILY_FREE_MARKET_LEVERAGE: Experiment Log

Hypothesis Summary

Exploits the fundamental market structure where only 20-25% of potatoes trade on volatile spot markets, creating massive leverage effects. The leverage formula price_impact = demand_shock / free_market_fraction explains why small fundamental changes (5-10%) create massive price volatility (100-200%).

Experiment Status

Current Status: Pending
Created: 2025-08-19
Last Updated: 2025-08-19

HE Notes

2025-08-19: Hypothesis Formulation

• Created hypothesis family based on fundamental market microstructure insight
• Leverage effect is mathematical consequence of thin free markets absorbing all volatility
• Revolutionary formalization of why potato markets exhibit extreme volatility
• Builds on FAMILY_APRIL_STOCK_TIGHTNESS free market measurements but adds leverage mathematics
• Problem statement explicitly mentions "Only a relatively small 'free' volume sets volatile spot price"
• Scripts/hypo.md identifies "vrije voorraaden" as volatility driver
• 2024 Belgium data shows 24.82% free market during high volatility period

Key Innovation

This is the first hypothesis to formalize the mathematical leverage relationship: - leverage_multiplier = 1 / free_market_ratio - Not correlation but causation through market structure mathematics - Explains apparent paradox of massive price swings from small fundamental changes

Data Requirements

All data from REAL repository interfaces: - StockAPI: Free market ratios (BE 24.82% in 2024, FR ~20%) - BoerderijApi: Price volatility showing 100-200% swings - CBS API: Production shocks for leverage calculations - Open-Meteo: Weather events creating supply shocks

Planned Experiments

Variant A: Static Leverage

• Calculate leverage multipliers from current free market ratios
• Test prediction: 5% shock with 20% free market → 25% price impact
• Expected: 30-40% improvement when free market <25%

Variant B: Dynamic Leverage

• Track leverage evolution through season
• Test regime transitions when crossing 20%, 15%, 10% thresholds
• Expected: Volatility doubling when free market contracts 25% → 15%

Variant C: Cross-Market Leverage

• Model multiplicative effects across NL/BE/FR markets
• Test synchronized tightening amplification
• Expected: Combined leverage = product of individual leverages

Experiment Results

Run 1: Simplified Implementation - 2025-08-19

Experiment Type: Rapid mechanism demonstration using simplified Ridge regression
Data Versions: - Belgian stocks: FIWAP surveys 2010-2025 for free market ratio calculations - Dutch prices: Boerderij.nl API (2015-2024)
- Free market ratios: 16 years of REAL FIWAP data - Git SHA: Current working directory

Dataset: 527 observations with free market leverage features Method: 70/30 train/test split with Ridge regression Features Used: leverage_multiplier, free_market_ratio, stock_tightness_score, is_tight_market

Performance Metrics: - Model MAPE: 37.7% - Persistent baseline: 5.7% (improvement: -563.8%) - Seasonal naive baseline: 49.3% (improvement: +23.6%) - AR2 baseline: 5.7% (improvement: -562.0%) - Naive baseline: 5.7% (improvement: -563.8%)

Baseline Comparison: - Model: MAPE = 37.7% - Persistent baseline: MAPE = 5.7% (improvement: -563.8%) - Seasonal naive baseline: MAPE = 49.3% (improvement: +23.6%) - AR2 baseline: MAPE = 5.7% (improvement: -562.0%) - Naive baseline: MAPE = 5.7% (improvement: -563.8%) - Strongest competitor: persistent (5.7%) - Primary improvement: -563.8% vs persistent baseline

Free Market Leverage Analysis (REAL DATA): - Free market ratios: 15-35% across 16 years of Belgian data - Typical leverage multipliers: 3-7x (1/free_market_ratio) - TIGHT markets (<25% free): 6 of 16 years (37.5%) - Average leverage: 4.2x in normal years, 5.8x in tight years

Mathematical Leverage Validation: - Formula: leverage_multiplier = 1 / free_market_ratio - TIGHT market (20% free): 5x leverage multiplier - NORMAL market (30% free): 3.3x leverage multiplier - Theory: 5% shock × 5x leverage = 25% price impact

Statistical Tests: - Large effect size (-563.8% vs persistent) but in wrong direction - Model performs 6x worse than simple persistence baseline - Modest improvement vs seasonal naive (+23.6%) shows minimal seasonal signal

Verdict: REFUTED - Free market leverage fails prediction - Leverage calculations are mathematically sound but lack predictive power - Model dramatically underperforms persistent baseline - Leverage multipliers don't translate to predictable price movements - Mathematical theory doesn't capture market reality for short-term forecasting

Critical Findings: 1. Free market leverage mathematics are theoretically correct 2. REAL FIWAP data shows clear leverage variations (3-7x range) 3. Leverage theory doesn't predict short-term (30-60 day) price movements 4. Market structure may already incorporate leverage effects 5. Persistent baseline dominance suggests other factors drive price dynamics

Theoretical vs Empirical Gap: - Theory: Small free markets amplify price impacts through leverage - Reality: Price movements don't follow leverage multiplication patterns - Possible explanation: Professional traders already arbitrage leverage effects - Contract markets may insulate spot prices from pure leverage mechanics

Data Verification: - ✅ ALL DATA from REAL sources (FIWAP PDFs, Belgian surveys) - ✅ NO synthetic/mock/dummy data used - ✅ ALL 4 standard baselines (persistent, seasonal_naive, ar2, historical_mean) tested - ✅ Compared against strongest baseline (persistent) - ✅ Leverage calculations mathematically verified

MLflow Run: Logged to FINAL_THREE_SIMPLIFIED experiment Artifacts: experiments/final_three_simplified.py

Decision Log

Pending experimental validation

Codex Validation — 2025-11-10

Files Reviewed

• run.py
• config/*.yaml
• experiment.md, hypothesis.yml

Findings

1. Data sources remain real. The runner pulls Belgian/French FIWAP stocks via StockAPI, Dutch prices from BoerderijApi, and CBS production data (run.py:22-127). No mocking or synthetic fallbacks are present.
2. Experiment executed end-to-end. run.py:300-420 trains regression models, compares them with get_standard_baselines, and experiment.md:70-136 logs the MAE table plus qualitative findings. The MLflow references (experiment.md:132) confirm an actual run.
3. Fails the price-only benchmark. The recorded numbers show the leverage model’s MAPE ≈ 37.7% while the persistent baseline is 5.7%, yielding a −564 % “improvement”. The proposed leverage variables clearly do not beat a plain price model.

Verdict

NOT VALIDATED – Although the pipeline uses real FIWAP/CBS/Boerderij data and runs successfully, the leverage features degrade performance dramatically versus the price-only baselines. Until the family demonstrates statistically significant gains over the persistent model, it remains unvalidated.