The AFML Division of Labour: Why One Quant Overfits

The lone-quant problem

When one person sources data, engineers features, designs the strategy and runs the backtest, the same incentives that produce a strategy also judge it. Selection bias and unrecorded trials follow almost inevitably.

Separating the stations

The proposed alternative is a research assembly line: data curators, feature analysts, strategists, independent backtesters, deployment and portfolio oversight. Each station has a defined input, a defined output and no incentive to flatter upstream work.

The backtester in particular should not see the strategist's motivation, only the specification, so the evaluation is adversarial rather than confirmatory.

How Voltacent Agent applies it

Voltacent Agent runs a Head of Research coordinator and six specialist roles as separate model calls. Each stage receives only the upstream artifacts it is allowed to see, so the backtester stays independent of the strategist.