FAMILY_AUTOML_PATTERN_DISCOVERY - Experiment Log

Overview

Objective: Use AutoML frameworks to automatically discover hidden patterns and beat the current 53.7% improvement baseline at 12-week horizons through computational brute force and automated feature engineering.

Status: DEVELOPMENT
Priority: CRITICAL
Target Performance: 60%+ improvement (vs current 53.7% baseline)

Variants

• Variant A: AutoGluon comprehensive ensemble with 1000+ automated features
• Variant B: H2O AutoML with deep learning and gradient boosting focus
• Variant C: TPOT genetic programming for evolutionary pattern discovery

Experimental Plan

Data Requirements

• Primary: BoerderijApi weekly Dutch potato prices (1,203 observations)
• Secondary: NDVI satellite data, international markets, weather accumulation, storage indicators
• Validation: USE ONLY REAL DATA from repository interfaces - NO SYNTHETIC DATA
• Horizon: 12-week (84 days) - proven optimal for maximum improvement

Feature Engineering Strategy

1. Automated Feature Explosion: Generate 1000+ features programmatically
2. Price transformations: lags 1-52, moving averages, differences, ratios
3. Polynomial interactions: 2nd and 3rd order combinations
4. Fourier transforms: seasonal cycles at multiple frequencies
5. Wavelet decomposition: multi-scale temporal analysis
6. Technical indicators: RSI, MACD, Bollinger Bands

7. Cross-market features: international spreads, correlations

8. AutoML Framework Testing:

9. AutoGluon: 2-hour comprehensive ensemble search
10. H2O AutoML: 2-hour deep learning + GBM focus

11. TPOT: 4-hour genetic programming evolution

12. Validation Protocol:

13. Rolling-origin cross-validation with corrected baselines
14. Test against ALL 4 standard baselines: persistent, seasonal_naive, ar2, historical_mean
15. Statistical testing: DM + HLN, TOST vs SESOI, FDR correction
16. Compare against STRONGEST baseline (lowest error)

Success Criteria

• Primary: Beat 53.7% baseline by 5%+ relative improvement (58.7%+ total)
• Statistical: p < 0.05 vs strongest baseline after FDR correction
• Practical: Improvement exceeds 8% SESOI threshold
• Discovery: Identify top 20 predictive features humans wouldn't consider

Experiment Results

Results will be appended here as variants are completed

Decision Log

Decision summary will be added after all variants complete