Model A · product
Risk signal
0.6339
ROC-AUC fold-mean OOS · carx h=1 · anchor
Probability of a large abnormal move / cascade. Position gate passed. Not a buy/sell recommendation.
📊 Public accuracy statistics
We publish our own performance metrics so you can judge whether to trust the system.
Live feed — high-signal events severity≥6: 59,506; total archive: 148,039; latest: 8/6/2026, 3:46:53 AM
Last real metric recalc: 7/29/2026, 11:12:36 AM
Model B · audit
Direction
≈0.50
ROC-AUC OOS · CAR sign · four configs
Measured near chance. Not offered as investment advice until skill and legal status are proven.
Public figures use the open corpus. A tighter estimate on your private sources is possible only in the member cabinet and only after a separate calibration on your contour — public numbers are never copied over. Request access →
Loading OCI…
Verified retro cases (2024–2026)
Public 2nd–3rd order cascade timelines from external sources. Oracle feed cross-check — honest gap, not “we predicted T₀”.
- C1 · Red Sea shipping · third-order supply chainORDER-LAG
- T₀
- 2023-11-19
- First Oracle hit
- 2025-01-27
- Lag
- ~14 mo
- Gap
- ORDER-LAG
T₀ missed; Oracle first captured rerouting narrative ~14 months late. Dashboard severity 6 without order tag.
- C2 · Sanctions / tech decoupling cascadeORDER-LAG
- T₀
- 2024-06-12
- First Oracle hit
- 2025-06-04
- Lag
- ~12 mo
- Gap
- ORDER-LAG
EU/US tech sanctions T₀ absent in DB; Jun 2025 geopolitical cluster only. RegTech thesis zone, not live feature.
- C3 · USDRUB / CBR key rate macro cascadeLOW SEVERITY
- T₀
- 2024-08-14
- First Oracle hit
- 2025-06-04
- Lag
- ~10 mo
- Gap
- LOW SEVERITY
CBR Aug 2024 hike missed; Jun 2025 CBR ops at severity 5 — below dashboard threshold. Ticker live, event layer quiet.
Not investment advice. Severity and probabilities are model estimates. Oracle ingest does not cover calendar 2024; cases use external timeline + partial DB match.
📈 Model calibration
Resolved outcomes (30d)
1866
positive class: 0
Brier Score (30d)
0.267
0 = perfect, 0.25 = random
ROC-AUC (90d)
0.454
1.0 = perfect, 0.5 = random
Signals (90d)
9692
with outcome: 2206 / 9692
A hit rate on its own cannot be read without the base rate beside it. Where events of this class occur a quarter of the time anyway, a score of 28% is close to chance, yet it reads like a failing grade. The metrics published here carry their own random baseline instead: Brier and ROC-AUC.
Probability calibration (30 days)
Loading A/B…
How we measure accuracy
- Brier Score — mean squared error between predicted probability and actual outcome. Range 0–1; lower is better. A coin-flip baseline scores 0.25.
- ROC-AUC — area under the receiver-operating characteristic curve. 1.0 = perfect discrimination, 0.5 = random.
- Calibration — for predictions binned by probability, does the actual frequency match? E.g. 70% predictions should happen ~70% of the time.
The public site refreshes about every 5–10 minutes. Real-time feed and alerts live in the member cabinet. Calibration uses only signals with actual outcomes; the score does not “improve” without new outcomes. Misses downgrade graph edge weights.
Risk-signal calibration metrics cover only scenarios with a recorded outcome. Not investment advice and not a trade tip.
How the methodology works →
Checking access status…