When a product crack is already high (>75th percentile of its trailing five years), how likely is it to stay high — and does the answer change when inventories are scarce and refineries are maxed? Every number recomputes from the primary series; the method lives in ml/crack_persistence.py.
| regime at start (crack >75th) | n | +4w | +8w | +13w | +26w | median wks above p50 |
|---|---|---|---|---|---|---|
| TIGHT · inv ≤20 & util ≥80 | 68 | 90% | 70% | 55% | 45% | 11 |
| Low inventories only · inv ≤20 | 187 | 84% | 74% | 69% | 56% | 12 |
| Loose · inv ≥50 | 220 | 92% | 86% | 78% | 56% | 16 |
| All high-crack weeks | 565 | 89% | 81% | 74% | 57% | 15 |
1720 weeks scored, 1993-10-22 → 2026-10-02 · trailing-5y percentiles · start-weeks overlap, so read regime spreads, not decimals
| regime at start (crack >75th) | n | +4w | +8w | +13w | +26w | median wks above p50 |
|---|---|---|---|---|---|---|
| TIGHT · inv ≤20 & util ≥80 | 65 | 97% | 98% | 96% | 78% | 21 |
| Low inventories only · inv ≤20 | 230 | 92% | 82% | 79% | 71% | 21 |
| Loose · inv ≥50 | 208 | 95% | 90% | 87% | 76% | 57 |
| All high-crack weeks | 664 | 95% | 90% | 85% | 76% | 34 |
1720 weeks scored, 1993-10-22 → 2026-10-02 · trailing-5y percentiles · start-weeks overlap, so read regime spreads, not decimals
| regime at start (crack >75th) | n | +4w | +8w | +13w | +26w | median wks above p50 |
|---|---|---|---|---|---|---|
| TIGHT · inv ≤20 & util ≥80 | 28 | 89% | 96% | 96% | 57% | 20 |
| Low inventories only · inv ≤20 | 145 | 91% | 87% | 74% | 67% | 16 |
| Loose · inv ≥50 | 312 | 97% | 97% | 95% | 86% | 47 |
| All high-crack weeks | 647 | 96% | 93% | 89% | 79% | 39 |
1720 weeks scored, 1993-10-22 → 2026-10-02 · trailing-5y percentiles · start-weeks overlap, so read regime spreads, not decimals