How the model works
The monitor is a rule-based pipeline. Market observations become normalised features, features become weighted component scores, scores become a regime, and only then does the model produce a sentence.
The model asks where stress goes, not how bad markets feel
Financial stress rarely disappears when policymakers act on it. It is absorbed, it is displaced into another market, or it spreads. The monitor tracks which of those three is happening, in sequence, using rules rather than commentary.
Inputs are a fixed indicator registry
Fifteen daily and monthly series cover US yields and the curve, inflation breakevens, credit spreads, volatility, energy, gold, the dollar, USD/JPY and Japanese 10Y. Every indicator declares its stress direction, its frequency, a staleness rule and the correlation group it belongs to, so correlated measures are collapsed rather than counted twice.
Normalisation is robust, not naive
Levels are converted to comparable 0 to 100 scores using median and median absolute deviation over a rolling window, then a normal transformation. Velocity is change over five observations, acceleration is the change in velocity. Context indicators such as the dollar or USD/JPY score on movement, never on level.
Components are weighted, with weights renormalised for missing data
Structural resilience, system pressure, policy response, market confirmation, stress displacement, stress absorption, global transmission, acceleration and cliff risk each carry declared weights. When an input is missing or stale, its weight is redistributed and coverage is reported, so a partial score is never presented as a full one.
Intervention effectiveness is measured against the market, not the announcement
For each operation the model records the pre-intervention level, the trough that followed and the latest level. Effectiveness is the size of the initial improvement. Decay is the share of that improvement since surrendered. An operation can be large, well received and still be losing effectiveness.
Propagation requires sequence, not correlation
The chain from oil to inflation to US long-term rates to the dollar to Japan to global liquidity is directed. An edge only activates when the downstream market moves after the upstream peak and inside the expected lag window. Simultaneous moves do not count as transmission.
Regime classification is rule based
The regime, from green to critical, and the displacement classification are produced by explicit thresholds on the component scores and on confirmation breadth. Plain-English text is generated from the quantitative state. Text never determines the state.
Cliff risk is a proximity measure, not a forecast
Cliff risk combines pressure, confirmation and acceleration to estimate how close conditions sit to a region where small additional shocks produce disproportionately large responses. It is an initial calibration and requires validation. It is not a probability.
Data quality is published, never hidden
Coverage, freshness, breadth and sequence evidence produce a confidence band. Where feeds are delayed the monitor says so. Where a series blends frequencies, as the Japan 10Y history does, the caveat is stated on the details page.