Deep Reinforcement Learning with Function Properties in Mean Reversion Strategies

The paper explicitly derives (i) the per-timestep mean–variance objective and its maximizer E[δw_t]* = 1/κ with optimal reward 1/(2κ), (ii) a Bayesian prior P(a) ∝ exp(−c1·perr) and likelihood ∝ exp(c2(reward−optimal_reward)), yielding a log-posterior proportional to reward − (c1/c2)·perr, and (iii) with c1 = 10 and c2 = 2κ, the augmented reward R_t ≈ δw_t − (κ/2)(δw_t)^2 − (5/κ)·perr(t). The candidate solution matches each step. The only minor deviation is that the candidate informally suggests distributing perr across time in any manner summing to perr, whereas the paper defines perr(t) precisely via an averaging construction. Otherwise, the arguments and outcomes coincide.