AI Finds Pure Arbitrage in Options Markets and Builds Minimal-Risk Trades

When an options trade looks profitable, the hard part is knowing whether the edge is real or just hidden risk. This paper tackles that problem by directly spotting statistical arbitrage opportunities in options markets. The authors use a two-stage graph learning approach: first, they create a new prediction target that isolates pure arbitrage through synthetic bonds; then they train RNConv, a graph learning model with a tree structure, to predict it. They also design SLSA, a class of positions that represent pure arbitrage opportunities and are provably of minimal risk and neutral to all Black-Scholes risk factors under the arbitrage-free assumption. In tests on KOSPI 200 index options, RNConv statistically significantly outperformed graph learning baselines, and SLSA consistently produced positive returns. The average P&L-contract information ratio was 0.1627. The paper’s main contribution is a way to connect prediction and trading strategy so that model outputs can be turned into trades designed to capture these price gaps directly.

On KOSPI 200 index options, one new setup kept producing positive returns. Its average profit-and-loss per contract information ratio reached 0.1627. That number is not huge, but it is hard to fake with a clean trade. If you have ever wondered whether a winning move is real or just hidden risk, this work asks that same question. The key trick is to hunt for pure arbitrage, meaning a price gap with no obvious free lunch elsewhere. Then it turns that gap into a trade meant to stay neutral to the classic Black-Scholes model, the standard formula used to price options.

Why table-shaped data changed the game

Options data do not behave like a neat race chart. They look more like rows in a table. That matters because many graph learning models expect network-like links. Graph learning is a way to learn from links between things. Table-shaped features can confuse that setup. RNConv attacks that mismatch. It adds a tree structure to graph learning. A tree structure means a model that asks yes-or-no questions in layers, much like a decision tree. In tests on KOSPI 200 index options, RNConv beat the graph-learning baselines by a statistically significant margin. That means chance is an unlikely excuse. The same setup then fed SLSA, a position class meant to capture pure arbitrage. SLSA kept producing positive returns. Its average profit-and-loss per contract information ratio was 0.1627.

How RNConv and SLSA work together

The first stage builds a new target. Synthetic bonds, a made-up reference asset, help strip out the easy parts of the price move. That leaves a target aimed at pure arbitrage. RNConv then predicts that target. The model keeps the graph-learning frame. That frame learns from links between things. It also borrows a tree-like split rule from decision trees. That mix fits table-shaped inputs better than a plain deep network. The second stage turns predictions into SLSA positions. SLSA means synthetic long positions for arbitrage. Under the arbitrage-free assumption, which means no free lunch in the price system, those positions are provably of minimal risk. They are also neutral to all Black-Scholes risk factors, the usual option price drivers in that model. The SLSA projection performs that turn from prediction to position.

Why the split between signal and trade matters

This setup changes the shape of the problem. The goal is not only to predict a price gap. The goal is to predict a gap that can survive as a trade. That is why the new target matters. It filters the search toward pure arbitrage, instead of ordinary upside with hidden risk. It also explains the appeal of RNConv. A tree-based graph learner can handle table-shaped features without forcing them into a pure network mold. On KOSPI 200 index options, that split kept the edge tied to a trade. The result is not a promise that every market will behave the same way. It is a sign that prediction and action can be linked more tightly.

What to test next

The surprise now has a practical shape. A model can hunt the edge first. Then SLSA can turn that edge into a position meant to stay close to pure arbitrage. That is different from asking one model to do everything. It also fits the KOSPI 200 result. The 0.1627 information ratio is modest. It still shows that the split can work on market data. If this path holds, the prize is not a smarter guess. It is a cleaner trade.