Objective: This study examines the gap between the statistical loss function and the economic objective of stock selection, and assesses whether pairwise RankNet ranking—in a controlled comparison with pointwise mean squared error regression—improves cross-sectional selection quality and net portfolio performance. The roles of the financial loss function, hyperparameter optimization, and the components of the LSTM–RankNet architecture are also disentangled.

Methodology: The data comprise 46 symbols from the Iranian stock market over 2,618 trading days, from March 25, 2015 to February 25, 2026. The stock universe was not selected from an end-of-period list; at each formation date, the following criteria were applied using same-date information: a minimum tradable-days ratio of 80%, a maximum consecutive gap of 30 trading days, membership in the top two market-capitalization deciles, a minimum free float of 15%, and a 25% cap on the share of any single industry. The union of eligible symbols formed the parent universe, and a daily mask kept membership and tradability time-varying. Four objectives—MSE, RankNet, Sharpe-only, and Combined—were compared under a 0.4% transaction cost and ten shared seeds.

Findings: The average net Sharpe ratio of RankNet was 1.314, versus 1.108 for MSE. RankNet outperformed MSE in all ten seeds, and the 0.206-unit difference was significant at p = 0.002. The Combined model also outperformed both MSE and RankNet; however, Sharpe-only, with a Sharpe ratio of 3.819, surpassed Combined’s 2.784. In the single-run sensitivity analysis, removing HPO reduced the test Sharpe ratio from 2.750 to 1.030; yet, since this comparison was not replicated across independent seeds, the result is descriptive and does not suffice for generalizable inference about the HPO effect. The conservative DSR also did not fully rule out multiple-selection risk.

Originality/Value: The study’s contribution lies in providing a controlled comparison among pointwise, pairwise, and financial objectives in an emerging market, and in separating the “objective-alignment value” from the “architectural-complexity value.” The results indicate that the shift from pointwise prediction to decision-oriented learning is warranted, but that combining all deep components is not necessarily the optimal solution.