In this paper, we investigate a novel multiperiod portfolio decision model for loss-averse investors with dynamically adapted reference points in a market with serially correlated returns. We demonstrate that the optimal policy is a piecewise linear function of the deviation between current wealth and reference level, and its slopes are a path-dependent function of the historical returns, in sharp contrast to the constant slopes generated by the simplified model that ignores the diminishing sensitivity and assumes independent returns. We show that this new feature significantly changes the typical V-shape pattern of the risky position, resulting in a more complicated nonlinear functional mapping. Our research highlights the potential dangers of relying on the simplified model and provides valuable insights for investors and practitioners to develop effective portfolio strategies under realistic market conditions. Additionally, our simulation analysis indicates that the predictability of returns combined with a small degree of diminishing sensitivity may enhance the disposition effect. Lastly, we prove that the new policy also fits to solve the multiperiod mean-Conditional-Value-at-Risk (CVaR) portfolio optimization problem with correlated returns, further broadening the application of our findings.