The more factors, the merrier

Third in the series. Last time we showed how Edgelab values each instrument as written — every barrier, coupon and call — instead of flattening it into a straight line. But a portfolio is more than a list of instruments priced one by one. What hurts you is how they move together. This piece is about that: the vast universe of factors behind every number we produce, and how we capture the relationships between them at a scale most systems can't reach.

A portfolio's risk is not the sum of its parts. It's tempting to treat it that way — add up the risk of each holding, allow for some averaging between them, and call the result the risk of the whole. On an ordinary day, that's close enough.

On the days that actually hurt, it's badly wrong — because the averaging is the first thing to break. A portfolio can look well diversified and not be. Spread across regions, sectors and asset classes, it appears to stand on many independent legs; then a bad week arrives, the legs move as one, and the diversification that looked solid on the report turns out to have been a feature of calm markets rather than a property of the portfolio.

That's the problem this part of the engine exists to solve. Last time was about valuing each position faithfully on its own. But risk doesn't sit inside individual positions — it sits in how they move together, and in how that togetherness changes under stress. Capturing it means modelling not just a great many instruments but the full web of relationships between them, including the ones that only surface in a crisis. Most systems duck that problem at scale. We built for it. Here's how.

Diversification is not an average

Diversification is the closest thing in investing to a free lunch — but it's only as real as your grasp of how assets move together, and the usual tool for that is cruder than it looks.

Most models lean on correlation: a single number standing in for an entire relationship. It says almost nothing about the tails, which is the only place the relationship truly matters. Two assets can drift along independently for years and then fall in lockstep in a crash; one average buries that completely. Worse, the conventional well-behaved assumption understates how often things break down together in the first place — and breaking down together is the event that does the damage.

So we don't reduce the relationship to a number, and we don't assume its shape in advance. Instead we let market history show how things have actually moved together. The method has a name — filtered historical simulation — but the idea is plain: replay real history, including the violent episodes a tidy formula would smooth away, and read the relationships straight off what happened. The one catch is that a calm stretch from years ago and a turbulent week today aren't comparable, so each past move is filtered first — rescaled to current conditions before it's used. A quiet day from a sleepy year is scaled up to today's turbulence; a crisis day is put on today's footing.

History can only teach what it has already seen, and we're honest about that limit — it's one reason the scenarios built on top of this don't rely on the past alone, which is a story for later in the series. But what the process preserves is the full dependency structure — the entire shape of how things move with one another, tails included — rather than a single coarse summary. It's what lets the engine answer a question advisors are asked constantly and can rarely answer well: how much diversification do you actually have when markets are stressed, not how much you appear to have when they're calm.

The risk inside individual holdings

Broad market moves are only part of the story. Individual holdings carry risk that's specific to them, and it can jump for reasons that have nothing to do with the broad market.

Edgelab captures that idiosyncratic layer too, and can trace a single name's risk back to its cause — an incumbent whose risk rises, say, because a large technology player has just launched a competing product. The move is attributed to the thing that drove it rather than absorbed into a market average. When a position starts behaving differently from its neighbours, you can see why — at the level of the single instrument, not just the portfolio.

The limit most models run into

All of this depends on holding an enormous number of factors in view at once, and here's the obstacle the industry rarely advertises.

The textbook way to capture how factors move together is to build a correlation matrix: every factor lined up against every other. For a few hundred factors, fine. Push it toward hundreds of thousands and the matrix degenerates — it becomes mathematically unstable, and the relationships it reports stop carrying reliable meaning. The usual response is to shrink the problem: model fewer factors, herd instruments into broad buckets, keep the matrix small enough to behave. It runs. It also throws away the granular, name-level, tail-aware risk we've just described.

This is the corner most engines quietly cut — not for any lack of skill, but because the limit is real and they've chosen tractable over complete. It's the same fork we described in the first article, showing up again in a different place.

Preserving the relationship without compressing it

Edgelab takes a different approach — keep the dependency between assets while never explicitly building the correlation matrix.

We never assemble the fragile object in the first place. Because the relationships are read straight from the filtered history — where the joint moves of assets are already recorded as they actually happened — the dependence is preserved without ever being compressed into a giant table of pairwise numbers. The step that breaks at scale is simply skipped.

That one decision is what lets the universe stay enormous — more than a quarter of a million risk drivers — without collapsing into noise. It's the difference between modelling the whole world coarsely and modelling a large, detailed part of it faithfully. And it's why "the more factors, the merrier" is meant literally here: more factors really do mean more of the truth, so we built the one thing that lets us carry them.

Robustness is built in

A universe this size can't depend on a person to hand-tune it whenever conditions shift. A model that needs constant recalibration can't run an entire book, every night, the way we described last time — roughly 1.5 billion calculations, finished before sunrise, the answer waiting in about half a second when you ask.

So we use a volatility model that needs no calibration. It adapts to fresh data on its own, and among the models that can run at this scale, its assumptions are the least contradicted by what the data actually does. More elaborate models exist on paper — but added complexity tends to buy fragility, and a model that errs more often at scale is no improvement. The honest aim isn't the most sophisticated model; it's the most accurate one that's also robust enough to run untouched.

And robustness here isn't a hope — it's enforced. Every nightly run passes through hard validation gates before a single number reaches you: limits on how much may fail to price, and on how far the results may drift from the night before. A run that doesn't clear the bar doesn't ship. The same rigour has been examined from the outside — EY reviewed the process independently for a global bank that's now a client, and J.P. Morgan put the methodology through its own vendor selection and chose it.

Carrying the whole universe is harder and costlier than carrying a convenient slice of it. We think that's the wrong place to economise, so we keep the complexity on our side of the line and hand you the clearer picture.