How a Compass analysis is built and checked.
Every number Compass shows you can be traced back through the steps below. The page is written for an investment-committee member who wants to know what to trust, how much, and why.
What a decision is
A decision is a named choice with money and a date attached: fund the plant, sign the offtake, hold the position, hedge the exposure. Compass records what you are deciding, the value you are deciding about, and the horizon by which the world will have answered. Everything else on the platform exists to put a distribution, not a point, on that value.
Where the scenarios come from
Each decision gets three to five scenarios: mutually exclusive stories about how the relevant drivers resolve by the horizon. The AI proposes a first set from the decision text, the asset class and the house views already in the library; the analyst edits, merges or pins them. A scenario is only kept if it names the drivers it moves and the evidence that would make it more or less likely, because that is what the next two steps consume.
Where the probabilities come from
Scenario probabilities come from an Analysis of Competing Hypotheses matrix: evidence items down the side, scenarios across the top, and a rating in each cell from strongly against (minus two) to strongly for (plus two), with a weight for how much the item should count. Each scenario’s score is the time-decayed weighted mean of its evidence ratings, scaled by a confidence factor; the probabilities are the softmax of those scores. Evidence weight halves every fourteen days, so a two-month-old headline cannot outvote last week’s data, and because the score is a mean rather than a sum, piling on correlated items does not manufacture certainty.
How drivers become distributions
Inside each scenario, every driver is entered as three numbers: a tenth percentile, a median and a ninetieth percentile. We fit a metalog distribution to those three points and fall back to a triangular distribution when the metalog is not feasible for that shape. Five thousand Monte Carlo draws per scenario turn the driver distributions into a value distribution, and the decision-level distribution is the mixture of the scenario distributions weighted by their probabilities. We report skew as quantile skew, the sum of the ninetieth and tenth percentiles minus twice the median, divided by the gap between the ninetieth and tenth, so a lopsided outcome reads the same way whatever the units.
How we score ourselves
When a scenario reaches its resolution date it is marked true or false and every author who put a probability on it is scored with the Brier score: the squared gap between the probability and the outcome, averaged over everything that author has resolved. Zero is perfect and a coin-flip forecaster scores a quarter on a 50/50 question. The reliability curve below groups resolved scenarios by stated probability and plots how often each group actually happened; a well-calibrated author sits on the diagonal.
What the AI does and does not do
The AI populates: it drafts scenarios, finds evidence, suggests the three driver percentiles from the data it can cite, and rates new evidence against each scenario. It proposes; it does not decide. Any cell, driver or probability an analyst pins stays pinned, and the AI’s later ratings are shown beside the pin rather than over it, so the record always says which numbers were the model’s and which were yours.
Data sources and cadences
Macro and rates series come from FRED and refresh daily. The chokepoint catalog (straits, pipelines, interconnects, single-source inputs) is maintained by hand and versioned. Supply chain edges come only from relationships cited in SEC filings, each reviewed before it is admitted; unreviewed machine-extracted claims never reach the graph. News signals are tagged hourly, and every number in the product carries its source and as-of date on hover.
The six-force model behind the coupling rules
Underneath the coupling rules that connect evidence to drivers sits a six-force model of how systemic stress moves: N (Narrative), G (Legitimacy), Q (Liquidity), S (Logistics), K (Lethality) and L (Learning). Each force is a node; the edges say which forces move which others, in what direction and with what lag. The model decides which drivers a piece of evidence is allowed to touch and how far a shock travels before it dampens. These letters appear only on this page; everywhere else in the product the force is named in words.
Deeper essays
The long form behind each step, for readers who want the derivations and the limits.