AI Daily Brief: The Backtest Was Perfect. The Market Disagreed.

The backtest was right. The market didn't get the memo.

The Quant Letter opened this week with something most quant teams already know but rarely say plainly: the gap between a machine learning strategy's backtested performance and what it does in a live market isn't random. It's systematic. That distinction changes what you need to do about it.

The mechanism is alpha decay, and it's been in the literature for a while. When a signal predicts something the market hasn't priced yet, trading it contributes to that pricing. If other strategies independently discovered the same signal, which with public ML methods and shared data sources they will, inflows toward the strategy erode its edge faster than any single team's in-sample results will show. The backtest worked because it wasn't competing with anyone.

A new paper formalizes this as "certified alpha capacity" thresholds: the maximum deployment size and time window before measurable decay, derived analytically rather than observed after the fact. That's a meaningful difference. Discovering your strategy stopped working in month four of live trading is one kind of problem. Being able to characterize beforehand exactly when and at what scale it degrades is a different kind. A companion paper, AlphaPareto, applies reinforcement learning to the same problem. Instead of optimizing for peak Sharpe on a single backtest window, it finds families of strategies that don't cannibalize each other's capacity constraints. That formulation is harder to build toward and considerably more honest about what you're actually trying to do.

Financial markets are not a stationary data-generating process. The act of learning from them changes them. That makes ML in trading structurally different from ML in almost every other domain, and it's why the backtest-live gap has a persistent component that no amount of hyperparameter tuning is going to fix.

A three-paper disagreement about vol surfaces (and who lost)

Two papers in the Journal of Futures Markets this year take different positions on where neural networks belong in options pricing. A third one stumbles into settling the argument in the least flattering way possible.

The first builds a hybrid: classical parametric model for the backbone, neural network correcting its residuals. On SSE 50 ETF options, the hybrid outperforms benchmarks on Sharpe and risk-adjusted returns. The second argues for going pure: a neural network learning the implied vol surface end-to-end, no parametric structure, with a meaningful RMSE reduction on American options. Both results hold up. The question is which architecture you trust when conditions shift.

A hybrid keeps interpretability. The parametric layer enforces no-arbitrage structure; the network handles empirical mismatch. A pure network can represent patterns the parametric model structurally can't, but you're relying on the training data and loss function to enforce no-arbitrage implicitly, which works until market conditions push into territory the training set didn't include.

Enter the third paper. Researchers trained a no-arbitrage neural network on liquid equity-market data and applied it, zero-shot, to a frontier market. Heston, recalibrated on sparse local data, beat it. The network that learned SPX vol surfaces doesn't transfer to Nifty options, because financial data doesn't generalize across markets the way image features generalize across photographs. The vol surface on an Indian index option has different liquidity conditions, different microstructure, different behavioral dynamics baked in. The network wasn't arriving somewhere familiar. It was working from a map of the wrong city.

Where neural networks have clearly won is speed. One commodity-options study reports inference at around 1.9 x 10^-4 seconds per option against Monte Carlo on the same grid, roughly five thousand times faster. When you're re-pricing a large book during an intraday vol spike, the difference between milliseconds and seconds is a real-time risk management capability, not a number in a table.

India: building what the Fed chose not to

SEBI's algorithmic trading framework, fully in force since April 2026, draws its line at ten orders per second and stamps every algorithmic order with an exchange-assigned Algo-ID. It's not a solution to the correlated-agent problem (what happens when strategies built on similar ML logic start triggering similar exit signals simultaneously during a stress event), but it's at least the audit infrastructure to investigate it afterward. The US Federal Reserve's SR 26-2 model risk guidance, released the same month, excluded generative and agentic AI from its formal scope entirely.

One regulator is building the paper trail. The other decided not to.