Sentiment Analysis

Sentiment systems operate at multiple levels of text resolution, each serving different trading horizons. Document-level analysis assigns a single score to an entire article or transcript, useful for macro event impact. Sentence-level analysis isolates specific statements, critical for parsing nuanced earnings call Q&A where a CEO's tone may shift between topics. Aspect-based sentiment goes further, extracting sentiment toward specific entities or attributes—for example, distinguishing positive sentiment toward 'iPhone sales' from negative sentiment toward 'supply chain constraints' within the same paragraph. Target-dependent models condition sentiment on the entity being discussed, preventing misattribution of a positive adjective to the wrong subject.

Two fundamental paradigms dominate sentiment system design. Lexicon-based methods rely on curated dictionaries like Harvard IV-4, Loughran-McDonald (finance-specific), or SentiWordNet, where words carry pre-assigned polarity scores. These are transparent and interpretable but struggle with context—'liability' is negative in general English but neutral in accounting. Machine learning approaches train classifiers on labeled corpora, learning contextual patterns. FinBERT, a BERT variant fine-tuned on SEC filings and earnings reports, captures domain-specific nuance that lexicons miss. Hybrid systems combine both, using lexicons for explainability and ML for accuracy. The Loughran-McDonald dictionary, with 2,709 negative and 354 positive words specifically calibrated for financial text, remains the gold standard baseline.

Sentiment signals exhibit rapid decay in predictive power. A positive news article's alpha-generating capacity may last seconds in high-frequency regimes or days in longer-horizon strategies. Systems must model sentiment half-life—the time for a signal's predictive value to halve. This requires timestamped sentiment scores with exponential weighting: recent sentiment carries higher weight in composite indicators. Event studies measure abnormal returns post-sentiment shock to calibrate decay curves. For earnings calls, sentiment extracted from the prepared remarks section decays differently than Q&A sentiment, as the latter often contains more spontaneous, information-rich language. Rolling window aggregation with configurable lookback periods allows strategies to tune sensitivity to sentiment velocity versus persistence.

General-purpose sentiment models fail catastrophically on financial text. Words like 'liability,' 'depreciation,' 'write-down,' or 'restructuring' carry negative connotations in everyday language but are neutral or procedural in finance. Domain adaptation techniques address this: fine-tuning on 10-K filings, earnings transcripts, and analyst reports; creating finance-specific stopword lists; and building custom polarity lexicons. The Loughran-McDonald Master Dictionary, developed by analyzing 10-K filings, reclassifies thousands of words that Harvard's dictionary miscategorized for finance. Contrastive learning on in-domain text pairs further sharpens model sensitivity. Without domain adaptation, a naive sentiment model might flag a company's 'aggressive depreciation schedule' as negative, missing that it signals conservative accounting and potential future earnings upside.

Global macro strategies require sentiment extraction across languages and markets. Cross-lingual transfer uses multilingual models like XLM-RoBERTa to project sentiment understanding from high-resource languages (English) to lower-resource ones. Translation-based pipelines convert foreign text to English before scoring, but introduce latency and potential meaning distortion. Language-specific fine-tuning on native financial corpora yields superior results but requires labeled data in each target language. Cultural differences in sentiment expression—Japanese earnings calls use more hedging language than American ones—demand calibration layers that normalize scores across markets. A 'cautiously optimistic' statement from a Tokyo-based CFO may carry different signal weight than the same phrase from a New York counterpart.

Sophisticated sentiment systems handle linguistic phenomena that defeat naive bag-of-words approaches. Negation detection identifies scope—'not profitable' flips polarity, but 'not only profitable' does not. Sarcasm and irony remain challenging; models must reconcile contradictory signals between literal text and context. Hedging language ('may,' 'potentially,' 'subject to') modulates sentiment intensity rather than direction. Contrastive conjunctions ('but,' 'however') signal sentiment shifts mid-sentence. Modern systems use dependency parsing to map grammatical relationships and attention visualization to verify that negation words attend to the correct sentiment-bearing terms. A system that misses 'We are not expecting a downturn' versus 'We are expecting a downturn' generates diametrically opposite signals from nearly identical text.