Absorption ratio and AR shift
The share of bank-return variance captured by the leading correlation-matrix eigenvalues, plus a short-minus-long moving-average shift in that share.
- Registry slug:
- absorption-ratio
- Visible surface:
- /systemic
Data and implementation
Vintage
2,951 absorption-ratio observations from 2020-04-03 and 1,130 AR-shift observations from 2021-12-31, both through 2026-07-06, measured 2026-08-30
Data tables
- data/parquet/systemic_series.parquet
Engine
- engine/finweave_engine/layers/systemic/absorption.py
Producer
scripts/build_systemic.py
Outputs
- data/parquet/systemic_series.parquet
Source: Yahoo Finance equity returns via the argus collector, internal-use input
Method
The engine computes a rolling correlation matrix for the fresh 27-bank panel, orders its eigenvalues, and divides the sum of the top 20% by total variance. The producer publishes this absorption ratio over both 63-day and 252-day windows.
AR shift is a separate series. It subtracts the 252-day moving average of the 252-day absorption ratio from its 21-day moving average, so it measures a change in coupling rather than the variance share itself.
Methodology evidence
| Path and lines | Evidence |
|---|---|
| engine/finweave_engine/layers/systemic/absorption.py:13-56 | Constructs the PCA variance share from the top 20% of correlation-matrix eigenvalues. |
| engine/finweave_engine/layers/systemic/absorption.py:77-98 | Constructs AR shift as the short-window average minus the long-window average. |
| docs/systemic_methodology.md:196-218 | Distinguishes the two rolling PCA windows from the separate AR-shift series and records their spans. |
| docs/systemic_methodology.md:269-338 | Records crisis anchors and independent recomputation validation for the systemic output. |
Equations
AR = sum(top eigenvalues) / sum(all eigenvalues)AR shift = short-window AR average minus long-window AR averageValidation
- The parquet contains 2,951 absorption-ratio observations from 2020-04-03 and 1,130 AR-shift observations from 2021-12-31, both ending 2026-07-06.
- The documented independent recomputation of the 63-day absorption ratio on 2023-03-17 matches the stored value at six decimal places (absolute difference 2.22e-16), and the COVID-period level is reported against the full-sample distribution.
Limitations
- The absorption ratio is a system-level co-movement statistic, not an institution-level loss, failure probability, or causal measure of fragility.
- The 63-day and 252-day absorption ratios are separate rolling constructions. Their rows are combined under one measure name and are distinguished by window_days.
- AR shift is available later because it requires 21-day and 252-day moving averages of the already rolling 252-day absorption series.
- The underlying Yahoo-derived equity-return input carries an internal-use, display-aggregates-only licence posture and is not a redistributable raw dataset.
References
- Kritzman, Li, Page, and Rigobon (2011), Principal Components as a Measure of Systemic Risk, Journal of Portfolio Management 37(4): 112-126.
Metadata endpoint
| Path | Method | Returns | Example |
|---|---|---|---|
| /api/methods/[slug] | GET | Registry metadata, implementation paths, measured vintage, methodology evidence, validation, and limitations. No model observations or parquet contents. | /api/methods/absorption-ratio |