Each cell of the State of Auto transition heatmap carries a number like +4.2σ. It says how far this month's rate sits from what was normal in a calm control period, measured in that period's own typical month-to-month spread. The control period is January 2017 to December 2019, the last stretch of ordinary years before the pandemic.
The formula
σ = (this month's rate − the control period's median) ÷ (1.4826 × the median distance from that median)
The median distance (the median absolute deviation) times 1.4826 is a standard deviation that one wild month cannot stretch: for well-behaved, bell-shaped data the two agree.
Nine made-up months
| Control month | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 |
|---|---|---|---|---|---|---|---|---|---|
| Rate | 21.0% | 22.0% | 22.5% | 22.0% | 23.0% | 24.0% | 24.0% | 23.5% | 28.0% |
| Distance from the median | 2.0 | 1.0 | 0.5 | 1.0 | 0.0 | 1.0 | 1.0 | 0.5 | 5.0 |
- In order: 21.0, 22.0, 22.0, 22.5, 23.0, 23.5, 24.0, 24.0, 28.0. The middle one, the median, is 23.0%.
- The distances in order: 0.0, 0.5, 0.5, 1.0, 1.0, 1.0, 1.0, 2.0, 5.0. Their median is 1.0 point; × 1.4826 = 1.48 points, one σ.
- This month is 29.2%: (29.2 − 23.0) ÷ 1.48 = +4.2σ.
With the mean (23.3%) and the standard deviation (2.0 points) instead, the one wild month (28.0%) stretches the spread and the same month scores only +2.9σ.
Reading the heatmap
- Within 1σ: no tint. 1σ to 2σ: a faint gold tint, worth a look. 2σ to 4σ, and beyond 4σ more strongly: green when the move is good, red when it is bad, gold where neither (a loan staying late).
- Up is good for pay-down and improve, bad for worsen and charge-off. +3σ on pay-down is good news; +3σ on charge-off is not.
- The number in the corner is the latest month alone; the line in the cell is a six-month average, and the shaded strip is the control period's middle 70% of months.
The real numbers
Prime auto loans (credit scores 660-900 at origination), the latest month (August to September 2026), against January 2017 to December 2019:
| Started the month | Pay-down | Improve | Stay | Worsen | Charge-off |
|---|---|---|---|---|---|
| Current | 4.3% (+2.8σ) | - | 93.6% (-2.9σ) | 2.1% (+0.4σ) | - |
| 1-29 days | 4.7% (+0.7σ) | 26.1% (-2.6σ) | 57.7% (+1.9σ) | 11.4% (+5.0σ) | 0.10% (+2.3σ) |
| 30-59 days | 3.6% (-1.5σ) | 32.5% (-3.2σ) | 40.0% (+3.1σ) | 23.5% (+1.9σ) | 0.49% (-2.3σ) |
| 60-89 days | 3.0% (-2.1σ) | 34.5% (-2.0σ) | 31.4% (+4.0σ) | 27.9% (-1.2σ) | 3.2% (-1.8σ) |
| 90+ days | 2.6% (-1.6σ) | 20.8% (-1.5σ) | 28.0% (+9.4σ) | - | 48.6% (-2.7σ) |
The 1-29 days → worsen cell: of the balance that started August 2026 1-29 days late, 11.4% was 30-59 days late or 60-89 days late or 90+ days late a month later. The control period's median was 7.5% and its σ 0.77 points: (11.4 − 7.5) ÷ 0.77 = +5.0σ.
Common pitfalls
- Mean and standard deviation. One unusual month inside the control period inflates a standard deviation and hides how unusual today is. The median and the median distance barely move.
- Unusual is not the same as bad. Read the direction: green and red say which.
- A different baseline gives a different σ. Against the last few years instead of 2017-2019, today's rates look less unusual, because those years were already high.
- One month. The σ is for the latest month alone; one odd month can flip a cell. Look at the line too.
Where to see it
- State of Auto deck (August 2026): the transition heatmaps on pages 46 (prime) and 81 (subprime); the full matrix and the scorecards (pages 47-52 and 82-87) use the same control period.
- The roll rates explainer, for how a transition rate is counted.
Source: LoanTape, SEC Form ABS-EE loan-level data, through September 2026.