About the application

This application implements the approach set out in my book, A Risk Manifesto. It reflects how I think material risk — which encompasses the unprecedented physical risks emerging in today's world — should be analyzed and managed.

It does not look at risk based on past return series. It looks at risk based on how the economy works, what companies do and depend on — products, facilities, supply chains, infrastructure, geographic concentrations — and works forward from there to what a disruption to any of it means for a portfolio.

The focus is on physical risk, structural risk, and material risk in a dynamic framework. Those terms, and the argument for approaching risk this way, are developed in the book. The application exists to make them concrete, as an illustrative and educational accompaniment to the arguments in A Risk Manifesto.

What it is for

The application answers questions at the core of risk decisions: What are you exposed to, and why? Where is that exposure concentrated? What connects holdings that appear unrelated? How would an event travel from the physical world through companies and into the portfolio? What is hidden by classifications built on historical market behavior? What are you missing?

Much of the analysis is qualitative. Material risks, the ones that reset the terrain rather than perturb it, are semantic before they are numerical. A firm's exposure to a Taiwan contingency is not a coefficient. It is a set of dependencies described in language: what it makes, what it needs, where it is made, what it cannot substitute. The exposure is stated before it shows up in prices.

Who it is for

The application is designed for portfolios that are diversified and conventionally structured — the sort held by an asset owner or long-term institutional investor. It is not designed for specialized and concentrated portfolios, and it will not attempt to analyze those.

Where it fits

In the years I spent as a chief risk officer, decisions were not made from a risk report. They start in a hallway or around a table. Someone names a worry, someone else reaches for an analogy, a mechanism gets proposed and argued over. It is after the question has taken shape that the risk metric machinery is put to the task.

This application is built to be the place you turn to have the conversation, naming the concern and questioning what it will touch. Pose the worry in the form you hold it — a blockade, a drought, a sanctions regime, a plant that stops — and the application works out what depends on what, how far the stress travels, and which holdings sit at the end of the chain.

What it does

The application traces how disruption moves along two channels. Firms are exposed because they depend on the same things and are connected by flows — a supplier, a corridor, a port, a grid. They are also exposed through similarity: when a category comes under suspicion, similar firms are repriced together, whether or not they touch. The first channel carries the physical disruption. The second carries the reaction to it.

Order matters as well as magnitude. The same set of events in a different sequence produces a different outcome, because each step changes the structure the next one moves through. If substitute capacity is built before demand shifts, the transition absorbs; if the shift arrives first, the shortage cascades. The output is dynamic, a path through the structure rather than a single exposure figure.

Scenarios start from the Base Scenario — the world as it stands today, with no shocks or events proposed — and depart from it. The scenario set is open. If you raise a risk I have not covered, I will look at whether the current data and models can support it. Some can be analyzed quickly. Others need relationships I have not yet mapped or data I do not yet hold; they may take several days to appear in the scenario list, or may not be possible to put there at all.

What the structural view adds is latent factors. These are dependencies that will move a group of firms together in often-surprising ways once conditions activate them, but that have never appeared in the past, so unseen through historical returns.

Standard factor output is reported for the Base Scenario, and not for the others. It remains the common language of the industry, and the contrast is useful: where the factor view and the structural view of a portfolio disagree, the disagreement is the finding. The reasoning is laid out in A Risk Manifesto and in the papers linked from the website.

How it works

The application is a conversation with a chief risk officer. You ask; you get a hallway answer — a paragraph or two, direct and complete for the moment, the kind of thing you get from someone whose time you are borrowing between meetings. You do not always have to ask. When something has changed, when a risk is emerging or amplifying, the CRO will raise it unprompted.

The analysis panel carries the evidence behind the answer: the affected holdings, the structural path, the current state of the scenario, and the sources. Links there are the CRO calling in the risk team for supporting work. The panel holds its place as the conversation moves on, and changes when a question calls for something else.

The division of labor is deliberate. The language model conducts the conversation. It phrases, connects, and judges what is worth raising. It does not compute the exhibits. Company exposures, product dependencies, latent factors, and every figure come from prepared scenario records and a deterministic compute layer. The model can tell you what the record says. It cannot invent what is in it.

When it comes to the data driving the application and model, I operate as a librarian, not as a data vendor. A data vendor says, "here is the answer." A librarian says, "this book should have what you need." Commercial data products ask you to trust the provider's collection, transformations, classifications, and corrections. Those might not be as well founded as users assume.

The application uses public material — government databases, regulatory filings, company disclosures, primary documents. Much of that material is text. For example, a 10-K names a production platform, a facility operating under a specific license, a sole supplier of a critical input, a distribution system that constrains what can be sold. It also states how those things connect: what depends on what, what enables what, what a change in one would foreclose. Language models make it possible to work in the medium the disclosures are written in, and to hold the relations between descriptions rather than replacing them with labels.

Source links are just getting started, but my objective is for every material assertion to trace back to where it came from.

For current developments I use a narrow set of sources: the Financial Times, the New York Times, Bloomberg, and the Wall Street Journal. If there is nothing in these, the concern is likely not material for a broad portfolio. I use nothing from social media, blogs, or general internet commentary.

These sources do substantive work. They tell me what the current environment is, what is tightening and where, which triggers are live, which bottlenecks are binding, whether the structure of a scenario has shifted, and how large a shock the situation now implies. A scenario is not a fixed object. Its state changes, and this reporting is how I track that and recalibrate. And it is what will occasionally lead the CRO to bring something up unprompted. I link to the original rather than reproducing it, which does mean some links sit behind subscriptions.

What it does not do