Laya-Finance

A fine-tuned version of Laya (ModernBERT-large decision encoder) for financial text decisions: sentiment, topic, central-bank stance, headline direction, finance terminology and trading-approach classification.

Derived from the original Laya (Apache-2.0). Fine-tuned on public financial datasets and a private corpus. This is the third public checkpoint; it replaces the second.

Intended use: educational and research purposes only. Not for commercial use. The restriction comes from the terms of some of the training data. The original Laya code and weights remain under Apache-2.0.

Usage

import laya

agent = laya.load("hravi/laya-finance")
q = {"type": "choice",
     "instructions": "What is the sentiment of this financial text toward the stock price or company outlook?",
     "criteria": {"bearish": "negative for the stock price or outlook",
                  "neutral": "no clear price impact, or purely factual",
                  "bullish": "positive for the stock price or outlook"}}
out = agent.predict_batch(["Operating profit rose to EUR 13.1 mn from EUR 8.7 mn."], {"a": q})
print(out[0]["answers"]["a"]["probabilities"])

Keep the criteria order fixed: it defines the label index. The model answers any multiple-choice question about a financial text, but it was tuned on the tasks below, and works best with the label sets used there.

Benchmarks

Held-out test accuracy, fp32, one forward pass per item. Laya-Finance runs at roughly 5 to 32 ms per item on an integrated GPU; the zero-shot LLM takes roughly 2 to 3 s per item.

Task	ProsusAI/finbert	Laya (original, zero-shot)	Laya-Finance	gpt-oss-20b (zero-shot)
Sentiment, PhraseBank	0.893	0.879	0.920	0.812
Sentiment, Twitter	0.725	0.767	0.895	0.740
Sentiment, FiQA	0.472	0.528	0.731	0.824
Topic	n/a	0.477	0.861	0.652
Headline direction	n/a	0.758	0.948	0.784
Central-bank stance	n/a	0.458	0.661	0.680
Finance terminology (domain)	n/a	0.604	0.877	0.834
Trading approach (style)	n/a	0.442	0.786	0.728

Notes:

• Laya-Finance beats FinBERT on all three sentiment sets. PhraseBank is the fair comparison, since FinBERT was trained on it. The Twitter and FiQA gains are partly from in-domain training data FinBERT never saw.
• The zero-shot LLM is clearly better on FiQA and slightly better on central-bank stance. On terminology and trading-approach macro-F1, Laya-Finance is level with or slightly ahead of it (0.810 vs 0.805 and 0.690 vs 0.670).
• Terminology and trading-approach test labels were produced and cross-checked by strong language models, not human annotators, so treat those two rows as indicative.
• FinBERT only supports sentiment, so other cells are marked n/a.

Using it as a first-pass gate

Laya-Finance is fast enough to screen every item, and its confidence is informative: the more confident it is, the more often it is right. A practical pattern is to accept confident answers and send the rest to a larger model or a person. Measured on the held-out test sets, ranking items from most to least confident:

Task	All items	Accuracy of the 80% most confident	Share of items answerable at 95% precision
Headline direction	0.948	0.991	100%
Sentiment, PhraseBank	0.920	0.972	90%
Sentiment, Twitter	0.895	0.957	83%
Topic	0.861	0.936	74%
Finance terminology	0.877	0.946	79%
Trading approach (style)	0.786	0.849	48%
Sentiment, FiQA	0.731	0.799	27%
Central-bank stance	0.661	0.717	18%

Read this as a best case: the confidence cut-off that gives each figure was chosen on the test set itself, so in practice choose it on held-out data. Confidence separates good from bad answers well for headline direction, short-text sentiment, topic and terminology, and poorly for FiQA and central-bank stance, where most items should go to a larger model. The combined pipeline with a larger model has not been evaluated here. Nuanced or long inputs (FiQA-style news, long documents, summaries, reasoning) are better handled by a larger model.

Limitations

• English text only; short passages (a few hundred words at most).
• Classifies text. It makes no claim to improve trading returns, and nothing here is financial advice.
• The zero-shot LLM is well ahead on FiQA (0.824 vs 0.731) and slightly ahead on central-bank stance.
• The rarest trading-approach classes have only 4 to 10 test items each, so their scores are noisy.
• Scores on the tasks above come from one training run, so differences of about one point are within noise.