Coherence training for forecasts · Kelvin, MATS
One week, every conditional
Manifold test events open on 2025-06-23 00:00 UTC. Qwen3-32B (no thinking) was asked each event alone, "A and B" and "B given A" both ways for every link, and "A and B and C" for every triangle. One joint model (event, link and triangle terms) was fitted to all 4,361 answers, exactly.
449events open
1030links
822triangles
15treewidth (exact)
0.501log loss, raw
0.474log loss, fitted
Stated vs fitted conditionals, largest group
The 16 strongest links. Stated: Qwen3-32B asked directly. Fitted: read off the joint model.
| A | B | P(B|A) stated | fitted | P(A|B) stated | fitted |
|---|---|---|---|---|---|
| Iran-Israel conflict Hormuz closure 2025 | Iran blocks Hormuz Strait in 2025 | 0.75 | 0.69 | 0.75 | 0.70 |
| Iran-Israel conflict Hormuz closure 2025 | Israel-Iran 2025: Hormuz Strait Closure | 0.85 | 0.75 | 0.75 | 0.65 |
| Israel-Iran 2025: Hormuz Strait Closure | Iran blocks Hormuz Strait in 2025 | 0.75 | 0.64 | 0.85 | 0.75 |
| Iran-Israel conflict Hormuz closure 2025 | Iran closes Hormuz Strait 2025 | 0.75 | 0.69 | 0.75 | 0.66 |
| Iran closes Hormuz Strait 2025 | Iran blocks Hormuz Strait in 2025 | 0.75 | 0.66 | 0.95 | 0.70 |
| Israel-Iran 2025: Hormuz Strait Closure | Iran closes Hormuz Strait 2025 | 0.75 | 0.65 | 0.75 | 0.71 |
| Strait of Hormuz Closure by 31 Dec 2025 | Iran-Israel conflict Hormuz closure 2025 | 0.85 | 0.60 | 0.85 | 0.73 |
| Iran-Israel conflict Hormuz closure 2025 | Israel-Iran 2025 Hormuz Closure <72h | 0.75 | 0.69 | 0.75 | 0.62 |
| Strait of Hormuz Closure by 31 Dec 2025 | Iran blocks Hormuz Strait in 2025 | 0.65 | 0.59 | 0.85 | 0.73 |
| Israel-Iran 2025 Hormuz Closure <72h | Iran blocks Hormuz Strait in 2025 | 0.75 | 0.61 | 0.35 | 0.68 |
| Strait of Hormuz Closure by 31 Dec 2025 | Iran closes Hormuz Strait 2025 | 0.75 | 0.61 | 0.85 | 0.70 |
| Israel-Iran 2025 Hormuz Closure <72h | Israel-Iran 2025: Hormuz Strait Closure | 0.75 | 0.67 | 0.85 | 0.64 |
| Strait of Hormuz Closure by 31 Dec 2025 | Israel-Iran 2025: Hormuz Strait Closure | 0.75 | 0.65 | 0.85 | 0.68 |
| Israel-Iran 2025 Hormuz Closure <72h | Iran closes Hormuz Strait 2025 | 0.45 | 0.62 | 0.65 | 0.66 |
| Strait of Hormuz Closure by 31 Dec 2025 | Israel-Iran 2025 Hormuz Closure <72h | 0.75 | 0.61 | 0.92 | 0.67 |
| Iran-Israel conflict Hormuz closure 2025 | Iran oil export halt to China 2025 | 0.75 | 0.67 | 0.45 | 0.48 |
Log loss against outcomes
| Question | n | Raw 32B | Fitted | Independent |
|---|---|---|---|---|
| Event alone | 449 | 0.634 | 0.693 | 0.634 |
| A and B | 1030 | 0.485 | 0.422 | 0.384 |
| A and B and C | 822 | 0.330 | 0.291 | 0.261 |
| B given A | 394 | 0.745 | 0.746 | 0.738 |
| All questions | 2695 | 0.501 | 0.474 | 0.440 |
Independent: the stated single-event answers multiplied. Conditionals scored where the condition came true. This week cost $1.53 (4,361 calls). All 35 test weeks: 231,293 calls, about $81.