A neural network is a flexible model that can approximate very complex relationships. That flexibility is its gift and its curse in markets.

What they add

Neural nets can capture non-linear, interacting patterns that a linear model or simple rule would miss — subtle combinations of volatility, flow, and structure. When there's genuine complex signal to find, they can find it.

The two costs

First, overfitting: with enough parameters a network can memorize almost any history, then fail the moment conditions shift. In a low-signal domain like markets, that failure mode is the default, not the exception (look-ahead bias). Second, the black-box problem: it's hard to know why a net made a call, which makes it hard to trust, debug, or bound with risk rules.

A model you can't explain is a model you can't govern. In trading, the ability to say “why did it do that” is a risk control, not a nicety.

How to use them sanely

The disciplined pattern is to use a neural net as one input feeding a transparent decision layer with hard risk limits — never as an unbounded autopilot. Guardrails, position sizing, and vetoes sit outside the net and can't be overridden by it. See why AI needs guardrails and why a raw model can't trade alone.