v21.0 tests whether the 8-symbol GOURMET system can predict cross-domain events by backtesting the 8 temporal windows against 90 days of historical data (March 28 – June 27, 2026). This transforms the system from descriptive to predictive.
- Do the 8 temporal windows (3, 4, 7, 55, 100, 111, 124, 127 days) predict cross-domain events?
- Which domain pairs show the strongest predictive correlation?
- What is the optimal prediction window for each domain pair?
- Does combining multiple windows improve prediction accuracy?
- Compile 90 days of cross-domain events from GOURMET grounding data
- Events categorized by domain: market, news, macro, entity
- Each event tagged with date, domain, magnitude, direction
- For each of the 8 windows, test: does elevated symbol score on day N predict an event in the next N days?
- Metrics: precision, recall, F1, accuracy
- Baseline: random prediction (50% for binary events)
¶ Phase 3: Domain-Pair Analysis
- Test all 12 domain pairs from v15.0 correlations
- Identify which pairs have predictive power and which don't
- Calculate optimal confidence thresholds per pair
- Test whether combining signals from multiple windows improves accuracy
- Example: 7-day + 55-day both elevated = stronger signal than either alone
- Use simple AND/OR logic, then test weighted combinations
- At least 2 windows achieve F1 > 0.65 (better than random)
- At least 3 domain pairs show significant predictive power
- Multi-window combination outperforms single-window prediction
- Results are statistically significant (p < 0.05)
- v21_0_backtest_results.md — Full results with metrics per window and domain pair
- v21_0_predictive_model.md — The predictive model specification based on results
- v21_0_validation_report.md — Independent validation against held-out data
- event_database.json — 90 days of categorized cross-domain events
- backtest_results.json — Structured metrics per window/pair/combination
- predictive_model.json — Model weights and thresholds
- v20.0: Initial predictive testing on 30 days showed 7-day window at 77.8%. v21.0 extends to 90 days.
- v19.0: Completed the 8-symbol system. v21.0 tests whether it predicts.
- v15.0: Identified 12 domain pairs. v21.0 tests which pairs have predictive power.
- v31.0: Knowledge graph embeddings. v21.0 provides the temporal prediction layer.
Status: Active
Vault Version: v21.0
Last Updated: 2026-06-27
Source Task: t_v21_1