Before It Trends: Weak Signal Detection in News Corpora for Strategic Scanning

The growing complexity of socioeconomic environments has increased the demand for tools capable of confirming early signals of change within a short window after their emergence, before they consolidate into observable trends. This study proposes a three-stage process for weak signal detection in news corpora that integrates Keyword Emergence Maps and Keyword Issue Maps (KEMs/KIMs), Transformer-based topic modeling, and time-series analysis, and specifies an evaluation protocol under which the contribution of each stage can be assessed. Applied to a corpus of 16,542 news articles spanning three years, the selection stages reduce a vocabulary of 14,672 terms to 423 candidates and then to six topics comprising between 0.14% and 1% of the corpus, so themes are surfaced by their rate of emergence while their volume is still negligible. The protocol comprises inclusion rules for the ground-truth event list fixed before any figure was computed, a peak-detection baseline that uses no event information, and a dispersion requirement for randomly initialized models. Under it, structural breaks align with documented events above arbitrary segmentation but not above the peak-detection baseline, and a recurrent forecaster does not improve on a random-walk reference. Detection one month after an event ranges from 36% to 82% across four detectors, so expected coverage depends on a choice that the evidence does not settle. These results locate the capability of the process in the construction of the topic series rather than in the temporal methods applied to it, and the process is proposed as a filter that reduces the volume of information that an analyst must review, with a measured confirmation delay, rather than as an autonomous early-warning system.