Peg / 03
checking…
Reading today
Fetching the stored result for the latest run.
tap a row for its historyloading
Where the models disagree p(depeg) · next session
USDT —
RF — LR —
USDC —
RF — LR —
DAI —
RF — LR —
0%scale to 40%40%
Waiting for both model outputs.
Deviation signed · vs peg
Relative stress arc = ±0.20%
−0.20
+0.20
Tolerance check box = ±0.10%
−0.20
+0.20
The Peg Report
daily analyst note · written 06:15 UTC
—

—

· checking
Retrieving the note and its verification record.
figures verifiedpending
unmatchedpending
written bypending
Tier 3
On what basis
Drivers —
Shap value · unitless · right pushes toward depeg
LSTM one-step fit
deviation % · peg 0
—
Loading the fit series.
What the model is actually predicting
No next-step forecast was returned for this asset.
The fit line is a one-step diagnostic on held-out data. A close fit is not evidence the model would see a depeg coming.
Tested on the event that matters —
—
Reported as counts, not a score. An F1 computed over two positive days would round to something confident and mean nothing; “caught 1 of 2, with 5 false alarms” is a claim a reader can check.
—
Loading the pipeline audit.
Tier 4 · historical
Behaviour under stress
Everything below is historical. None of it describes current conditions.
Behaviour when crypto falls —
—
Standard deviations are large relative to these means. Read the direction of the behaviour, not the precision of the figure.
Same test, traditional sell-offs —
—
Baseline · normal days for comparison
Value at risk —
AssetHist.Param.Cornish-F.
†
Reading this table
Historical is the empirical quantile; parametric assumes normality; Cornish-Fisher corrects for skew and kurtosis. Benchmarks are included deliberately — the comparison with stablecoins is part of the finding.
Event studies —
These panels read the raw per-asset frame, which keeps weekends, so the SVB break has always been drawn in full here. The classifier read the joined frame, which until this round did not — the same event down two paths, and only one of them was wrong. The holdout test in Tier 3 is scored on the corrected frame.
Tier 5
Provenance
Lineage
Prices ingested from the exchange API, cleaned and stored in Postgres with row-level security. The front end reads stored results only.
Schedule
One scheduled job at 06:15 UTC recomputes every figure on this page. Nothing is calculated in the browser.
Parity
Parity checks against the frozen dissertation code gate each release; a mismatch fails the build. Count pending — the suites have not recorded a run.
Fallback
Layered fallback so the page never shows a partial state. Layer names pending from the API.
Note author
Pending from the API.
Origin
Models and thresholds are frozen from the submitted MSc dissertation. Risk levels come from the API; the front end formats and never recomputes them.
Research output from an MSc dissertation on stablecoin stability. Not investment advice, not a recommendation, and not a guarantee of future behaviour.
Run 06:15 utc—