Loading...
Loading...
Contributors:
Dataset URI:
10.5281/zenodo.23150334Adaptive risk aversion in distributionally robust reinforcement learning is often implemented as an integrator that tracks twin-critic disagreement, and the increment channel of that integrator is rarely analyzed. We show that a proportional update has no bounded steady state: its expected displacement grows linearly with the number of updates at a rate set by the bias of its input, so the safety clip, not the data, fixes the risk parameter. The challenge is that the parameter is observed only through a noisy learned quantile estimate, so boundedness must hold uniformly over the horizon rather than on average. We state a channel condition for bounded displacement, zero gain of the increment channel at zero frequency, and propose a DC-blocking law, the gap minus its own exponential moving average, with a closed-form displacement and an endpoint dispersion independent of the horizon. Ten closed-form predictions were confirmed on synthetic gaps, and across 84 walk-forward runs on 34{,}571 five minute gold bars the proposed law held the risk parameter at $2.007 \pm 0.045$ without boundary contact, against $1.583 \pm 0.056$ for the proportional law; the six configurations span $1.9$ points against a minimum detectable effect of $0.6$ to $3.2$.
Keyword(s):
MeSH Term(s):
Grants and Funding:
No Grants or Funding Data Available
Associated Publication(s):
No Associated Publication Data Available
Supplementary Data:
No Supplementary Data Available
Related Dataset(s):
No Related Dataset Data Available
The National Library of Medicine (NLM) provides information about biomedical datasets. Inclusion on an NLM Web page does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: NLM Web Policies (nih.gov)