The news sentiment analysis module uses AI-powered sentiment analysis purpose-built for financial language — trained on SEC filings, earnings call transcripts, analyst reports, and financial news. Unlike general sentiment tools trained on product reviews, this AI was built specifically to understand financial language, where words carry domain-specific valence that generic models systematically misread.
The practical consequence is significant. Words like "headwinds," "cautious optimism," "maintaining guidance," and "in line with expectations" have specific positive or negative connotations in financial context that differ from everyday usage. A general model might score "maintaining guidance" as neutral or slightly positive. The AI scores it correctly as a soft negative in an environment where the market was expecting a raise. That distinction can meaningfully shift a ticker's sentiment sub-score in the Alpha composite.
Why general sentiment tools fail on financial text
The failure mode of general sentiment models on financial text is well-documented. Generic sentiment tools are significantly less accurate on financial text than purpose-built financial sentiment AI. The gap widens on earnings call transcripts, where the language is highly specialized: corporate euphemisms, analyst hedging language, and management rhetoric require domain-specific training to decode correctly.
The business consequence of misclassified sentiment is a corrupted Alpha sub-score. If a negative earnings call is scored as neutral, the sentiment input understates risk and the composite Alpha is artificially elevated. In a live trading context, that means a position might be taken at a moment when the smart money is already distributing. This domain-specific training is not a technical nicety — it is a direct risk management feature.
How the AI scores financial news
The sentiment AI outputs a confidence-weighted score on a positive-to-negative scale that feeds into the overall signal quality composite. The model is specifically trained to distinguish financial nuance — "maintaining guidance" scored correctly as a soft negative when the market expected a raise, versus a general model that would score it neutral.
Each article processed gets its own sentiment score. When multiple articles exist for a ticker in the lookback window, the system aggregates them with recency weighting — newer articles count more than older ones. An earnings call transcript processed early in the morning gets higher weight than a routine analyst note from three days prior. The aggregate sentiment score that feeds Alpha reflects the current news environment, not a trailing average that could lag actual developments.
How news sentiment is gathered
The full sentiment pipeline per cycle: fetch news from financial news providers for each candidate ticker in the scan universe (up to the most recent N hours of lookback), score each article for financial sentiment, aggregate scores with recency weighting, check for earnings within the next 5 days (which triggers a risk flag independently of sentiment score), and return a structured output containing the sentiment score, article count, and risk flags.
The critical safety rule: if the news API calls fail — if a provider is down, if the sentiment service returns an error, if the network call times out — the system defaults to blocking the trade, not clearing it. Not neutral. Not "assume no news." The system treats a failed news check as bad news. A trade will not proceed on a ticker where sentiment analysis could not complete. This is the conservative default that protects against executing into an undiscovered event.
How sentiment influences signals
The sentiment module's output contributes meaningfully to the Alpha composite, with its exact weight adjusted by the current regime classification. A sentiment score of +0.8 on a scale of -1 to +1 contributes a strong positive sub-score to Alpha. A sentiment score of -0.7 — which you would see after an earnings miss or a damaging regulatory ruling — can pull Alpha down significantly even if technical and forecast signals are positive. This is intentional: the system should be more reluctant to go long on a ticker that just received bad news.
Sentiment alone cannot push a signal above the Alpha floor. A perfectly positive sentiment score with weak signals from the other dimensions produces a modest Alpha, not a tradeable one. Conversely, a deeply negative sentiment score in an otherwise strong setup can drag Alpha below the floor, preventing a trade the technicals might have endorsed. The composite design means sentiment acts as a meaningful vote, not a veto — except in the most extreme cases where the halt-risk flag is set, which does block execution entirely.