This is a provocation. I'll admit that up front. No textbook calls Black-Scholes a graph neural network, no peer-reviewed paper makes the claim, and any quant reading the title is already preparing a counterargument. Good. Bear with me anyway, because the reframing, once you accept its looseness, does real work.

The core argument: every pricing model is a function mapping option inputs to a price. A graph neural network (GNN) maps a node's features and its neighbors' features to a representation. Classical pricing assumes each option is isolated, modeled with one node and no edges — effectively a GNN without neighbors. Modern graph-based pricing reintroduces the connections between options.

Strictly, calling Black-Scholes a one-neighbor GNN is a stretch, as a GNN with one node and no edges is not a GNN in any technical sense. But the analogy pushes us to reconsider a major modeling choice: what if assuming no neighbors is the field's most consequential simplification, overlooked because mathematics made it convenient?

The unstated assumption

Open any derivatives textbook and look at how an option is defined. Strike. Underlying. Tenor. Volatility. Risk-free rate. Maybe dividends. These are the features of the option as an isolated object. The pricing function operates on those features and returns a number.

This treatment isn't wrong, but it is a choice, and choices have consequences. The option being priced is, in fact, embedded in a dense network of relationships: other options on the same underlying at other strikes and tenors, and the same contract observed on prior trading days. None of this enters the pricing function. The model assumes it can compute a price from the option's own parameters alone, as if the option were truly an island.

In a frictionless, complete-market world with continuous trading, this assumption is internally consistent. The whole apparatus of risk-neutral pricing was built precisely to make this work — to construct a synthetic world in which the option could be priced from its parameters because the rest of the market had been collapsed into a single risk-neutral measure. It is elegant mathematics. It is also a heroic simplification.

Every quant who has worked with real market data knows where this simplification starts to leak. The implied volatility surface is shaped, not flat. Skew exists, and persists. The term structure exhibits features that no constant-volatility model can reproduce. Prices on the same contract move day to day in ways that a stateless pricer cannot see. The market keeps telling us that options do not live alone. The textbook keeps assuming they do.

Where the analogy starts to hold

A graph neural network (GNN) is a class of neural network models that operate on graphs, where data is represented as nodes and edges. In general, a GNN computes a representation for each node by aggregating information from its neighbors via a learned message function. The node's state is then updated based on this aggregated information, often through several layers. The overall output is typically a prediction about each node or the graph as a whole. Importantly, the graph's structure — how nodes are connected — is a primary input to the model.

Read that paragraph again with options markets in mind. The nodes are individual option contracts, each defined by strike, expiry, and type. The edges encode the relationships that matter for pricing:

The message function is whatever the model learns for each edge type — the way pricing information transmits from one contract to another along these three channels. Aggregation is how those channels combine to determine the price the node charges.

Framed this way, the relevant question isn't "is Black-Scholes a GNN?" but "what graph structure do options inhabit, and what would our models gain by seeing it?"

This modeling choice is foundational, and its convenience may have deterred deeper scrutiny.

What you gain by adding edges

This is where I should be careful, because it is also where my dissertation work lives. I am still in the middle of that research; the results are preliminary, and anyone who tells you GNN-based option pricing has been definitively solved is selling something. What I can say is what the structural argument suggests, with appropriate hedging on the empirical claims.

When you let a pricing model see the graph structure explicitly — when you give it access to the prices of neighboring options along the smile, along the term structure, and along its own recent history — several things change in principle:

None of this is magic. None of it is a free lunch. Each of these gains comes with costs in interpretability, computational expense, and theoretical tractability. Closed-form Greeks become numerical approximations. Hedging strategies that rely on closed-form derivatives have to be reconstructed. Black-Scholes is elegant precisely because it sacrifices structure for analytical convenience, and you cannot undo that sacrifice without paying for it somewhere.

What this is not

Three things this argument is not, so I can pre-empt the obvious objections:

This is not a claim that classical pricing is wrong. Black-Scholes, Heston, SABR, local-vol — all of these continue to do useful work, and probably always will. They are special cases of a broader picture, not refuted by it. The relationship is like Newtonian mechanics to relativity: still right within its domain of validity, illuminated rather than overturned by the more general framing.

This is not a claim that GNN-based pricing is currently better in practice. The empirical evidence is mixed, the methodology is still maturing, and the comparison to well-calibrated classical models is harder than enthusiasts often admit. In several settings I have studied, the structural model wins; in several others, the classical model wins. The interesting research question is not "which is better" but "where does each one work, and why."

This is not a claim that the three edge types I described are the only useful ones. Same-expiry, same-strike, and temporal edges are the ones my current framework uses. Cross-asset edges connecting options on correlated underlyings are a natural extension. Dealer-flow edges, regime-similarity edges, and macro-state edges are all plausibly worth exploring. Defining the right neighbor structure is itself a research problem, and the answer probably depends on the question you are asking.

Why this matters beyond pricing

The reason the framing is worth holding onto, even given its looseness, is that the same structural blind spot shows up across quantitative finance, not just in options pricing.

Credit risk models price each obligor as if its default were nearly independent of the rest of the portfolio, papering over the dependence with copulas that pretend to capture it after the fact. Risk parity allocations treat each asset as if its risk contribution could be measured in isolation. Factor models assume that each factor exposure can be priced solely as a function of its loadings. In each case, structure is present in the data, missing from the model, and shoved back in via correlation matrices that everyone admits are unstable and unsatisfying.

If the option-pricing argument generalizes — and I think it does — then the broader story of quantitative finance in the 2020s is not about better features or bigger models. It is about taking the relational structure of markets seriously enough to embed it in the model rather than treat it as an afterthought.

Where to be skeptical

If you have read this far and you are a working quant, you are probably already running the obvious counterarguments. Let me name a few, so we agree on what is being argued.

Classical models bake in structure, too. The local-vol model has a state-space structure; Heston has a stochastic-vol structure; and copulas have a joint structure. The argument is not that classical models are structureless. It is that the structure in classical models is parametric and pre-committed, while the graph-based structure is learned and can be richer than any prior parameterization allows. Different kinds of structure, with different tradeoffs.

Calibration to observed prices already captures market structure indirectly. True. But it captures it through the prices of each option in isolation, not through the relationships between options. A well-calibrated Heston model can reproduce the observed implied vol surface; it cannot tell you why the surface has the shape it does, and it cannot extrapolate to surface points it has not seen with the same fidelity as a model with explicit neighbor structure can.

Black-Scholes cannot be a GNN with a single neighbor, because a GNN with a single node has no neighbors to pass messages to. Correct. The framing is a useful provocation, not a formal claim. Read it as "classical pricing models are the degenerate case you get when you assume no graph structure exists," and the argument holds without straining the technical definition.

The next decade

The way I read the state of quantitative finance is that the field is at a transition point similar to where computer vision was in roughly 2010. The dominant paradigm — hand-crafted features, parametric models, and careful calibration — is mature, well-understood, and increasingly limited. A newer paradigm learns the representation from structure, letting the data tell you what matters, which is technically possible, empirically inconsistent in early results, and politically contested. The field will spend the next several years figuring out which problems are best solved by which approaches, and the answers will not be clear-cut.

The honest position for a researcher right now, I think, is that classical pricing models will continue to anchor the field for the foreseeable future, and graph-based approaches will progressively colonize the places where the classical assumptions break down hardest: exotic structures, illiquid markets, regime transitions, and cross-asset effects. Not a replacement. An extension.

Black-Scholes is a graph neural network with one neighbor. Not literally. But the looseness of the analogy is exactly the point. Once you see what a different choice of graph could unlock, the old textbook treatment stops looking like a definitive answer and starts looking like a default assumption that has not yet been seriously questioned.

It is being questioned now. That is what makes this an interesting decade to be doing this work.