← Risk Co-Pilot Free preview

How every figure is measured

Every number this tool reports, the method that produced it, and what it will not support. If something here is wrong, write and say so — [email protected]. The whole product rests on this page being true, and one person reads that address.

The window

Every figure is measured over the stretch where all of your holdings have prices — the days where any one of them is missing are dropped for the whole portfolio, so that every holding is compared over the same days.

The consequence is worth stating plainly, because it surprises people: the youngest holding sets the start. Add a fund launched eighteen months ago to a book you have held for twenty years, and every figure on the page is measured over eighteen months. The window is printed beside the reading everywhere it appears, for exactly this reason.

The prices

End-of-day closing prices from public sources. No intraday data, no quotes, nothing streamed: this tool measures what a portfolio has done, and the close is the right resolution for that.

Holdings quoted in another currency are converted into your base currency at each day's rate, so the reading is what you actually experienced rather than what the asset did locally. The gap between those two is itself reported, as currency exposure — and the hedged leg there is a perfect, costless hedge, which a real one is not.

A ticker that will not load is named, never silently dropped. A portfolio tool that quietly analyses four of your five holdings is worse than one that refuses to answer.

Every figure

The code beside each name is the number the report cites it by, so a report you have copied and this page can be read side by side. Read off the engine at build time, by running it over six example portfolios. Several figures are conditional and appear only where they mean something, so this is the union of what those portfolios produced, not a fixed list.

F1Holdings

count of positions

F2Effective bets

Meucci effective number of bets on a Marchenko-Pastur denoised covariance, PCA basis

Describes how many independent directions the risk occupies. It is not a forecast of loss and does not predict a drawdown.

F3Risk in the largest shared direction

variance share of the leading principal component

This is commonality, not a fault. A diversified equity book is MORE driven by one shared direction than a single stock is, because diversifying is what removes the idiosyncratic part.

F4Annual return

compounded, monthly rebalancing

Past window only. Says nothing about the next one.

F5Volatility

annualised standard deviation

F6Return per unit of risk

Sharpe ratio, computed on returns in excess of T-bills

F7Worst fall on record

largest peak-to-trough decline inside the window

The worst fall *in this window*. A longer window would very likely contain a worse one.

F8Holdings leading that direction

largest absolute loadings on the leading component

Loading signs are fixed so the largest is positive; a negative loading means that holding moves against the others.

F11–F15Each holding: money and risk

Euler risk contribution on the same denoised covariance as F2

The ratio is the finding. A weight alone cannot say whether a position is large, because that depends on everything else held.

F20Risk, earlier stretch then later stretch

the last 252 trading days against the 252 before them

F21What changed the risk

exact additive split of the change in variance into individual volatilities and co-movement, averaged over both orderings (Shapley)

Today's weights are held fixed across both stretches. This is what the MARKET did to the book you hold now — not a history of the book, which this app does not store. Reading it as 'my trades caused this' would be exactly backwards.

F22Average pairwise correlation

mean of the off-diagonal denoised correlation matrix, each stretch

F2BEffective bets, measured the other way

Meucci ENB in a minimum-torsion basis rather than the PCA one

The two bases disagree here, so part of this reading is the lens and not the portfolio. Take the count as a range, not a point.

What is not computed

No forecast of returns, losses, prices, or the probability of any of them. No recommendation to buy, sell, or hold, and no suggested allocation.

That is not modesty. This project pre-registered and then abandoned 27 predictive hypotheses — regime models, conditional distributions, quantile forecasts, stress persistence — each one tested against a threshold written down before the data was seen. The page listing them is public because a claim to describe rather than predict is only worth reading if it cost something.

What is known about you

Your portfolio. It is kept in your own browser and nowhere else. There is no account and no sign-up. Holdings are sent to the server to be measured and are not stored there — which also means that if you clear your browser data, they are gone, and we cannot restore them.

The counting. Page views are counted, in aggregate. Each day a random secret is generated, held in memory and never written down; a visitor's address is hashed with it to make a short identifier that distinguishes one reader from another for that day only. The secret is discarded when the day rolls, so yesterday's identifiers cannot be matched to today's or reversed to an address. What is saved is the count, not the identifiers. No cookies, no analytics vendor, no third-party script anywhere on this site.

You can read the same numbers. They are at /stats.

What this is

A free, non-commercial preview, built and run by one person. Nothing is charged, nothing is sold, no price data is redistributed or offered for download. It is early: expect rough edges, and expect figures to be added rather than removed. Questions, corrections and complaints all go to the same place: [email protected].

Descriptive portfolio analytics for research and education. Not investment advice; no recommendations and no price predictions. Read a finished analysis at /examples, or measure your own at the app.