Developing a model of companies' investment attractiveness based on the "news sentiment" approach
TL;DR
The evidence suggests that timestamp integrity, the informational concentration of the headline and lead, abstention, and exit design may matter more than encoder complexity.
Abstract
The study develops and empirically evaluates a method for assessing companies' investment attractiveness by converting corporate news into measurable signals for the analytical module of a personal stock-exchange trading bot. The design integrates event-study logic, ticker-timestamp identification, intraday price windows, market-adjusted returns, and selective classification with an abstention option. The empirical corpus contains 213,533 news records associated with 498 companies from a static end-2025 500 constituent list. For each event, the pipeline constructs standardized fixed windows and an adaptive window ending at the first crossing of a 20-period weighted moving average. TF-IDF, BERT, FinBERT, FinBERT-tone, and DistilFinRoBERTa are compared across four text configurations under a 0.60 confidence threshold. Mean accuracy is 0.425, macro-F1 is 0.347, and the Matthews correlation coefficient is 0.056; hence, conventional classification statistics do not support a claim of strong multiclass prediction. Economically relevant separation emerges only after confidence filtering and transaction-cost adjustment. TF-IDF produces the best net result in 22 of 37 horizon–window cells and wins all seven adaptive-window comparisons. The highest net direction-corrected cumulative-return sum is 7.166 for the one-hour WMA-break specification, closely followed by 7.160 at 30 minutes. The evidence suggests that timestamp integrity, the informational concentration of the headline and lead, abstention, and exit design may matter more than encoder complexity. The reported magnitudes are comparative research statistics rather than estimates of production-ready portfolio profitability.