Goodhart’s Law
Goodhart’s Law is the principle that “when a measure becomes a target, it ceases to be a good measure” — once an objective is optimised against a proxy metric, agents (or optimisers) find ways to raise the proxy without advancing the true goal. It appears in this vault as the conceptual root of Reward Specification Error and Reward Design Sensitivity: a Markov Decision Process Trading Model optimises whatever reward it is given, so a reward that is only a proxy for genuine risk-adjusted profit will be “Goodharted” — the backtest PnL curve rises while true tradeable edge does not. It also underlies Data-Snooping Bias, where a backtest performance figure used as a selection target stops measuring out-of-sample skill once strategies are tuned to maximise it.
Connections
- Reward Specification Error — relates, optimising a proxy reward Goodharts the true objective, source: https://arxiv.org/abs/2201.03544
- Reward Design Sensitivity — relates, tuning a reward to maximise a reported metric is Goodharting that metric, source: https://arxiv.org/abs/2201.03544
- Data-Snooping Bias — relates, a backtest metric used as a selection target stops measuring genuine skill, source: https://arxiv.org/abs/2201.03544