Reinterpreting Short-Term Prediction Signals as a Post-Signal Realization Diagnostic: Evidence from KOSPI
YeoHoon Yoon, Kyung Sung Kim · The Korean Data Analysis Society · 2026
This study moves beyond the conventional practice of evaluating daily top-decile prediction signals on stocks listed on the Korea Composite Stock Price Index (KOSPI) by whether they are correct at a single future date, and instead reinterprets them from the perspective of a market movement that appears shortly after the signal. Conventional machine-learning studies in finance focus on whether a prediction score is correct at a fixed single horizon, yet upward and downward phases often build up gradually, so the information in a signal may be realized soon after the signal rather than exactly on the labeled date. A distinctive feature of this study is that this shift in perspective is offered not as a new model but as an evaluation framework for the temporal realization patterns of already-generated prediction scores. Using daily KOSPI data, we select the highest-scoring stocks each day and examine, across a broad range of return targets, multiple model families, and two feature settings, whether the predicted movement is realized within a short follow-up window, with a one-year gap between training and testing to limit label overlap. The signals appear more closely linked to a movement occurring soon after the signal than to a correct prediction on the exact target date, and this pattern holds consistently across a range of checks—including window-length and top-quantile sensitivity, panel-aware bootstrap, and year-by-year stability—and is more pronounced for downward movements. These results suggest that short-term prediction scores can serve as practical indicators for screening and monitoring possible near-future movements rather than as classifiers tied to a single target date. The study contributes a practical framework for evaluating prediction signals through their temporal realization patterns.