polibert/swik-sentiment-labels

swik Financial Sentiment Labels

Asset-specific financial sentiment labels for 35 securities — commodities, FX, indices, and crypto.

v0.2 — Updated 2026-04-04. 50,589 total records.

What makes this different

Standard financial sentiment datasets assign generic polarity to headlines. This dataset applies asset-specific inversion context from the swik inversion catalog — a community-maintained knowledge base of how phrases actually move prices for each specific asset.

Example: "OPEC cuts production" → generic NLP scores negative. For crude oil (OIL), it's bullish — less supply means higher prices.

Schema

Column	Description
text	Headline or news snippet
label	bullish / bearish / neutral / irrelevant
label_id	0=bullish, 1=bearish, 2=neutral, 3=irrelevant
security	Asset symbol (OIL, GOLD, BTC, EURUSD, ...)
magnitude	Signal strength 0–1
relevance	How relevant headline is to this asset 0–1
source	human or ai

Dataset Stats (v0.2)

Split	Rows
train	40,471
test	10,118
total	50,589

Label	Count	%
bullish	28,580	56.5%
bearish	15,318	30.3%
neutral	6,505	12.9%
irrelevant	186	0.4%

Top securities by coverage:

Symbol	Records
OIL	16,737
LNG	4,632
BTC	3,758
BRENT	3,342
ETH	3,307
EURUSD	3,141
NIKKEI	2,080
NATGAS	1,989
NDX	1,425
RBOB	1,217

Label Sources

• human — community contributors via swik.io/contribute/label
• ai — generated by Claude Haiku (claude-haiku-4-5) with swik inversion catalog injected

Transparency note: The majority of labels are AI-generated using the swik catalog as context. Human labels are ongoing. AI labels use asset-specific inversion patterns to correct for well-known domain inversions (e.g. supply cuts → bullish for commodities). Filter by source == 'human' for the highest-quality subset.

Benchmark Results

Evaluated against the matched inference corpus (community_labels_legacy, 30,053 pairs):

Model	Accuracy	F1 macro	Notes
heuristic-v1	98.88%	0.981	Keyword-based, deterministic
haiku-4-5 (baseline prompt)	39.6%	0.309	LLM, confidence anti-calibrated
haiku-4-5 (variant B, calibrated)	46.0%	0.456	Improved prompt under testing

Note: benchmarks measured against AI-generated labels. Human-label benchmark pending.

Related

• Model: polibert/swik-heuristic-v1 — keyword baseline
• Platform: swik.io

License

CC BY 4.0 — free to use. Cite swik.

Citation

@dataset{swik_sentiment_labels_2026,
  title={swik Financial Sentiment Labels},
  author={swik Community},
  year={2026},
  url={https://huggingface.co/datasets/polibert/swik-sentiment-labels},
  license={CC BY 4.0}
}