Approach
We publish what we believe. We do not publish how to rebuild it.
The framework below is what a client is entitled to understand before they rely on us. The construction is proprietary, and stays that way.
Risk has more than one face
A single measure of market risk is always wrong somewhere. Credit spreads miss a political rupture. Sentiment surveys miss a funding squeeze. Volatility is the last to know.
GXPOL reads macro risk through a set of independent lenses — financial conditions, real-economy behaviour, political and policy pressure, positioning, funding and liquidity, the physical economy, corporate solvency, cross-border coercion, and the structural cost of transition. Each is built separately, from its own data, and scores the same question from its own angle.
Each lens is meant to use its own inputs, so the composite does not count the same driver twice. A check runs on every code change and blocks any new overlap. A short list of older overlaps is still being worked through, and it is listed in the methodology paper.
Not everything moves together
This is the part most composite indices get wrong, and it is the centre of our approach.
Some lenses measure stress that is happening now. Others measure conditions that precede stress — and those frequently move the opposite way during a crisis. Central banks ease. Curves steepen. Commodity demand collapses. Averaging everything into one number blurs precisely the signal you need.
So we do not. We maintain a read of the regime you are in, and a separate read of stress occurring right now. Their agreement is confirmation. Their disagreement is the warning.
One number is not a signal
An elevated reading on its own is a poor basis for doing anything. Across two decades of history, our headline measure sits above its risk threshold on roughly two days in five — far too often to act on.
So a reading has to be corroborated before it becomes a call. We raise WATCH when the composite is elevated and at least one independent lens agrees. We escalate to ALERT when global stress confirms it as well as a domestic one. Out-of-sample, corroboration costs us nothing in detection — it misses exactly the events the raw reading misses — and takes the false-alarm rate from roughly 39% of calm days to 15% at WATCH and 3% at ALERT.
Both stages are timestamped and recorded when they happen, with the lenses that corroborated each one. A call you cannot date is not a track record.
Discipline
Evidence, including when it is inconvenient.
-
Three optimisations we rejected
We have now tried three times to improve on equal weighting: a velocity blend, volatility-scaled thresholds, and per-economy optimised weights. Walk-forward testing rejected all three. The per-economy weights were the closest — and at matched false-alarm rates they bought one extra detection in twenty-nine, because they simply made the index more sensitive, which a threshold does more honestly. So we ship equal weights. Most vendors would have fitted them, reported the in-sample figure, and said nothing. The evidence is in the client methodology paper.
-
Historical readings use only what was known at the time
A historical reading uses only what was knowable on that date. This is harder than it sounds and easy to get quietly wrong — a model that rewrites its own past through today's lens will backtest beautifully and fail you live.
-
What happens when a source fails
Inputs stop updating and sources change. When a source fails, its part of the score is dropped and the remaining weights are rescaled; the gap is not filled with a neutral value. Readings show how much of their weight is backed by live data, and we are extending that figure to every engine.
-
How the record works
Each weekly reading is published with a timestamp and a hash of the conditions that produced it. If the methodology changes, history can be recalculated under the new version; the published record is kept, and the change and its effect are written down.
What we will show you
Clients and serious prospects receive the full methodology paper: the lens architecture, the independence constraints, the regime thresholds and how they were calibrated, the out-of-sample testing, and — plainly stated — where our coverage is weaker and why.
It is enough to satisfy a quantitative due diligence. It is not a blueprint, and it is not published here.