📈 finvader

Finance-aware sentiment analysis for Rust

VADER, re-tuned for the market — financial lexicon, phrase rules, and catalyst detection for news headlines and market text instead of tweets.

🎯 Why finvader?

Generic VADER was calibrated for social media. On financial text it misfires in two directions:

❌ Generic VADER problem	💥 Example	finvader fix
Misses finance sentiment	"beats expectations", "cuts guidance", "going concern" → score ≈ 0	📖 Financial lexicon + phrase rules
Misfires on neutral finance words	"gross* margin", "cancer drug", "debt refinancing"* → scored negative	🎭 Neutral-override masking

And it adds something VADER never had: 🚨 catalyst detection — single events (FDA approval, buyout offer, index inclusion, bankruptcy) that permanently re-rate a stock and deserve to be surfaced on their own.

🏆 Accuracy

On a hand-labeled set of 60 financial headlines:

Analyzer	Correct signal	Accuracy
generic VADER	31 / 60	51.7%	🟥🟥🟥🟥🟥⬜⬜⬜⬜⬜
finvader	60 / 60	100%	🟩🟩🟩🟩🟩🟩🟩🟩🟩🟩

🚀 Quick start

[dependencies]
finvader = "0.1"

use finvader::{FinVader, Signal};

let fv = FinVader::new();
let s = fv.analyze("Acme beats Q3 expectations and raises full-year guidance");

assert!(s.compound > 0.5);
assert_eq!(s.signal, Signal::StronglyBullish);

// Which terms drove the score:
for t in &s.triggers {
    println!("{:<20} {:+.3}", t.term, t.valence);
}

💡 Construct one FinVader and reuse it across calls — it loads the lexicons once and is Send + Sync.

⚙️ How it works

Each input runs two parallel passes — a masked base-VADER pass and a financial layer (phrases, gap-phrases, single words) — then the two are blended, nudged by any catalyst, and clamped:

Stage	What it does
🔡 normalize	Lowercase, strip punctuation, keep market-text characters (`-`, `%`, `$`, `'`)
🎭 mask_for_base	Replace finance-neutral words (`gross`, `cancer`, `debt`, `crude`, `vice`, `share`…) with a placeholder before base VADER, so everyday valences never pollute the score
🧩 phrase / gap / word passes	Match multi-word phrases (`beats expectations`), gap-tolerant pairs (`beats … expectations`, up to 2 tokens apart), and single words — consumed so nothing double-counts. `up 45%` / `down 30%` become magnitude-scaled moves
🔁 negation & boosters	`failed to beat expectations` flips · `sharply missed` amplifies · `slightly missed` softens · `beat by 40%` scales on magnitude
⚖️ blend	With financial terms present: `0.35 × base + 0.65 × financial`; otherwise base VADER passes through untouched
🚨 catalyst bonus	A detected event shifts the compound by ±0.25

📊 Signals

compound maps to a discrete Signal:

compound	Signal
`>= 0.5`	`StronglyBullish`	🟢🟢
`>= 0.15`	`Bullish`	🟢
`-0.15 .. 0.15`	`Neutral`	⚪
`<= -0.15`	`Bearish`	🔴
`<= -0.5`	`StronglyBearish`	🔴🔴

🚨 Catalyst detection

Beyond the smooth sentiment score, finvader flags episodic-pivot events and returns them as Option<Catalyst>:

📈 Bullish catalysts	📉 Bearish catalysts
FDA approval / clearance	FDA rejection
Breakthrough therapy	Missed / failed endpoint
Met primary endpoint	Chapter 11 / Chapter 7
Buyout / takeover / merger	Going concern
S&P 500 inclusion	SEC charges
Contract awards	Accounting fraud
Record quarter, beat-and-raise	Auditor resignation

🔌 Where it fits

finvader is the scoring core of a news-driven momentum alert pipeline:

⚡ Performance

Single-threaded, release build, cargo run --release --example bench (Apple Silicon):

Analyzer	Per headline	Throughput
generic VADER	1.8 µs	~570,000 / sec
finvader	22.3 µs	~45,000 / sec

finvader does more work per call (masking + three match passes + catalyst detection) and still clears tens of thousands of headlines per second per core.

📖 Lexicon

Single-word and phrase valences are calibrated for market news, informed by the Loughran-McDonald financial sentiment research. Valences are on VADER's -4.0 ..= 4.0 scale.

🧪 Examples

cargo run --example demo              # side-by-side finvader vs generic VADER
cargo run --release --example bench   # throughput benchmark

finvader 0.1.0