Interchange: change here between the Markets and ML lines.
RealHedge: does a simulator-trained hedger survive the real market
Why
Papers on neural options hedging almost always report their advantage measured inside the same simulator the model was trained on. That advantage might not survive contact with real market paths and the spread a trader actually pays. I wanted a number for the gap, not an assumption about it, tested against a classical band rule that has been in use since 1997.
How
Before any model was trained, I wrote down the exact predictions and the exact test that would confirm or kill each one, and sealed a two-year block of real price history by its file hash so it could only be looked at once. The comparison runs a neural hedger against three classical baselines on real SPY and TSLA price paths, at a range of trading costs, and the metric is the size of the worst outcomes, not the average one. Every correction made along the way, including two the plan itself had wrong at the start, is dated and kept in the record rather than edited away.
What came out
Nothing to report yet. Trains stop here from the opening date.
What broke
The first draft of the claim had a sign error: it predicted the neural hedger's edge would reverse below a certain trading cost, when the trading theory says the opposite direction, caught before any model was trained. The primary trading cost was initially copied from a prior paper at double the correct value, because that paper states its rate per round trip, not per trade. And a full pull of two-year daily closing prices found the ordinary vendor closing price disagreeing with the exchange's own official closing print by up to 287 points on some days, on prices that were already being used, which changed which price the whole study now trades at.
Mind the gap
Private for now. No hedging result exists yet, development or holdout. It goes public with a licensed repository once there is one to report.