<![CDATA[Predicting stock price movements remains challenging due to nonlinear, dynamic market behavior. News sentiment and technical indicators (RSI, moving averages) each show predictive value, but few studies integrate both within a single, robustly validated framework. This study proposes an algorithmic trading framework integrating daily news sentiment from four pre-trained transformer models (ProsusAI/finbert, nlptown/bert-base-multilingual-uncased-sentiment, yiyanghkust/finbert-tone, and soleimanian/financial-roberta-large-sentiment) with RSI and moving average filters to generate buy signals on 30 NASDAQ-100 stocks. Beyond standard backtesting, the framework is validated through five analyses: benchmark comparison, exit-rule sensitivity, look-ahead-bias alignment, statistical testing, and walk-forward validation. Over a 2-year simulation (57,434 news articles), buy signals were triggered when RSI fell below 40, price pulled back 0.5–5% below MA20, and price remained within 10% above MA50, with entry on day i+1 and a profit-triggered exit. All four models produced positive cumulative returns (123.33–187.84%, Sharpe Ratio 2.00–2.73), remaining profitable under transaction costs up to 0.3% per trade. Benchmarking showed roughly half the Maximum Drawdown of naive baselines despite lower raw returns, favoring selectivity over return maximization. The stop-loss/take-profit variant substantially reduced performance, while stricter news-timestamp alignment left results robust (Sharpe Ratio above 1.5). Statistical tests found three of six model-pair differences significant at the 5% level, though the two top-performing models are statistically indistinguishable (p = 0.546). Out-of-sample evaluation confirmed profitability in the most recent test segment, with acknowledged sub-period variability. These findings position news sentiment as an effective, risk-disciplined complement to technical analysis, with model selection depending on investor priorities.]]>
