Historical-Memory Equity Selection: A Comparative Study of Retrieval, Neural Prediction, and Adaptive Mixing

Motivated by recent progress in long-context sequence modelling, this study examines whether modern forecasting methods can support intermediate-horizon equity selection and whether a bounded numerical memory can make their evidence more inspectable. We implement a memory-centered framework in which forecasts are constructed from eligible historical records, their completed 63-session returns, and an explicit aggregation rule; the memory can operate independently or alongside neural predictors. The evaluation compares 16 policies across 103 equities in six markets, represented by 204 continuous market/run accounts over 1,568 union sessions from 2020 through 2025. Similarity memory averages the outcomes of 25 precedents selected by Euclidean distance between standardized, six-session-pooled annual feature windows. Controls include random eligible precedents, an unconditional historical mean, plain nearest neighbours, ridge and neural predictors, momentum, and passive equal weight. Passive equal weight has the highest mean-market Sharpe (0.636) and six-year annualized return (11.91%). Similarity memory records Sharpe 0.245 and annualized return 2.99%, below random memory (0.423) and the historical prior (0.352). All six prespecified memory contrasts are negative, and neither these comparisons nor the two gate-versus-mixture contrasts reject at the 5% level after Holm adjustment. A complete six-fold execution generated more than four million policy predictions; its summaries report full required-forecast coverage and reproduce the primary analysis. The framework makes the retrieval calculation decomposable, but the multi-year evidence does not show that similarity selection improves portfolio performance. JEL Classification: C45 , C52 , C58 , G11 , G17