Funding professionals are acquainted with this downside. A backtest could present a sample. An element could seem important. A mannequin could determine construction in returns. But as soon as turnover, buying and selling prices, slippage, market influence, regime shifts and operational constraints are included, the obvious alternative could disappear.
AI doesn’t get rid of this downside. In some instances, it will possibly intensify it.
A reinforcement studying agent is educated to behave. If it detects weak alerts, it might attempt to exploit them by means of frequent place modifications. However when these alerts are near noise, buying and selling exercise can grow to be pricey. The agent could overfit short-lived fluctuations, interact in extreme buying and selling and erode returns by means of prices.
Alternatively, the agent could be taught the alternative lesson. If the surroundings is noisy and transaction prices are significant, essentially the most rational coverage could also be to do nothing. On this case, the AI system doesn’t grow to be a superior dealer. It turns into inactive.
Each outcomes are necessary. They present that failure is just not random. It’s informative.
If an AI agent can’t remodel a detectable sample into optimistic internet efficiency, that will inform us one thing concerning the market surroundings. The sign could also be too weak. The associated fee construction could also be too excessive. The statement house could also be inadequate. The mannequin could also be too constrained. Or the sample could also be statistically actual however economically untradable.
That is the essence of the Learnability Threshold.
