Methodology · validation
The validation record
This is the only page on the site where a claim about how well this model has performed is permitted to
appear. Nothing on the read, the monitor or the posture map carries a track record, because an accuracy claim
standing next to a live number reads as a forecast of that number.
Basis of every result below
Every entry below is a retrospective simulation. The engine did not publish an edition at the time of these episodes; the scores were computed later, on data as it stands today rather than as it stood then. Under R13 that inflates any accuracy claim by whatever the subsequent revisions were worth. Nothing here is a live record, and the two are never merged into one number.
Page dated 25 August 2026 · method version
asymmetric-investor-brief:2026-08 · results are not merged across bases, and a live
record cannot begin until per-observation vintage fields are published.
Closed episodes
Definition · sample · vintage · method
Each episode states what was being predicted, over what sample, on which data vintage, by which method — and
what the result was. The disconfirming-evidence field is mandatory and cannot be published empty. Four
retrospective hits on four selected episodes is a sample of four, chosen after the fact.
Global financial crisis
Retrospective simulation2008-09-01 → 2008-10-31
- Event definition
- Failure of a top-five US broker-dealer followed by money-market fund breaking par.
- Sample and exclusions
- US equities, EM equities, US high yield. Excluded: private credit (no comparable series in 2008) and crypto (did not exist).
- Data vintage
- Latest available, not point-in-time
- Method version
- v2.0 applied retrospectively
- What the model said
- System average −0.79
- What happened
- S&P 500 −57% peak to trough; EM −65%
Disconfirming evidence — mandatory fieldThe score reached −0.79 only after Lehman filed. On 2008-08-31 the same computation returns −0.31, which would not have cleared the CRITICAL threshold. The model reads the crisis, it does not lead it.
COVID crash
Retrospective simulation2020-02-19 → 2020-03-23
- Event definition
- WHO pandemic declaration plus co-ordinated national lockdowns.
- Sample and exclusions
- US equities, credit, oil. Excluded: EM (policy responses too heterogeneous to score as one).
- Data vintage
- Latest available, not point-in-time
- Method version
- v2.0 applied retrospectively
- What the model said
- System average −0.70
- What happened
- S&P 500 −34% in 33 days
Disconfirming evidence — mandatory fieldAn exogenous biological shock is not a macro-regime signal. The framework had no inflation, credit or valuation reading that anticipated it, and a −0.70 print two weeks in is a description rather than a forecast. Counting this as a hit flatters the engine.
Bond and duration repricing
Retrospective simulation2022-01-01 → 2022-10-31
- Event definition
- Fastest policy-rate normalisation since 1994 with negative stock-bond correlation.
- Sample and exclusions
- Global aggregate bonds, US equities, crypto. Excluded: commodities (war-driven, separately caused).
- Data vintage
- Latest available, not point-in-time
- Method version
- v2.0 applied retrospectively
- What the model said
- System average −0.74
- What happened
- Global bonds −20%; crypto −65%
Disconfirming evidence — mandatory fieldThis is the one episode where the framework's inflation and rates inputs should have led, and the retrospective score only clears CRITICAL in Q2. It also gave no signal on the October turn, so a reader following it would have been late out and late back.
Regional banking stress
Retrospective simulation2023-03-08 → 2023-03-31
- Event definition
- Uninsured-deposit run at a top-20 US bank followed by two further failures.
- Sample and exclusions
- US regional banks, US equities, credit. Excluded: European banks (separate idiosyncratic case).
- Data vintage
- Latest available, not point-in-time
- Method version
- v2.0 applied retrospectively
- What the model said
- System average −0.60
- What happened
- US banking index −28%
Disconfirming evidence — mandatory fieldThe broad index recovered within weeks and the S&P ended the quarter higher, so a −0.60 systemic print substantially over-read the event. This is a false-positive-shaped hit: right about banks, wrong about the system.
Regime audit
Where the model and its own output disagree
Separate from episode scoring: an audit of whether what the site publishes is coherent with what its own
components read. Unresolved rows stay on the page as unresolved.
- Published regime, issue No. 7
- AMBER — stagflation, high conviction
- Implied by the current board
- AMBER — anchors stretched, tactical benign
- Agreement
- Agrees on level, disagrees on breadth
- Unresolved
- HIGH conviction is published while the three classes disagree. Under K3 the honest label is closer to BIFURCATED. Carried as an open audit item rather than silently corrected.
K10 kept: the engine scores its own published regime against what the computation now implies, dated, every edition.
Source tiers, and one breach
The rule and its exception
Our own hierarchy admits tiers one to three for a published flag. One live indicator rests on a tier-five press
source. It stays on the page with the breach marked, and it is excluded from the coverage aggregate — reporting a
higher coverage figure by quietly counting it would be the worse failure.
Where tiers conflict, the higher tier stands and the conflict is published. A T5 source may provide colour but is never used to establish a flag. One tripwire currently breaches this rule and is marked.
Limits of this record
Read before quoting any figure above
US-centric by construction
Every admitted tripwire is a US series. Non-US assessments are inferred from US conditions plus central-bank divergence, not measured locally.
Weekly cadence
The engine runs once a week. Intra-week discontinuities are invisible until the next edition, and the fastest scenarios are precisely the ones that resolve inside a week.
No cross-asset correlation model
Cells are scored independently. The model cannot tell you what happens when three of them move at once, which is what a cascade is.
Private-credit opacity
Marks are model-derived, amendments suppress defaults, and the disclosure cadence is quarterly at best. The private-credit readings are the least reliable on the board.
Intervention override
A central bank or treasury can invalidate a threshold overnight. The board measures conditions, not the reaction function.
No sovereign directional signal in systemic stress
In severe system-wide stress, reserve-currency sovereign bonds may rally as a flight-to-quality beneficiary regardless of what the rates readings say. The engine goes blind on sovereigns in exactly the scenario a reader most wants a sovereign view.
What would make this a live record
- Per-observation vintage fields — observed_through, first_published, revision_state — on every indicator, so a score can be reconstructed as it stood on the day.
- A published composite for every weekly cycle, including cycles where the composite did not move, so the sample is every week rather than every interesting week.
- A pre-registered definition of what counts as a hit, dated before the episode it scores.
- A corrections register, so a revised input is visible as a revision rather than as a different number.
Thresholds, indicator definitions, the blind-spot rules and the version history live on the methodology page.
The indicator readings themselves live on the monitor, with their sources one click away.