FLAME2 — Financial Language Analysis for Multilingual Economics v2
One model. Ten languages. 150,000 headlines. Perspective-aware financial sentiment.
FLAME2 is a multilingual financial sentiment classifier that labels news headlines as Negative, Neutral, or Positive — but unlike other models, it does this from the local investor's perspective of each economy.
The same news can mean opposite things for different markets:
- "Oil prices fall to $65/barrel" → Negative for Arab markets (oil exporter) / Positive for India (oil importer)
- "Yen weakens to 155 per dollar" → Positive for Japan (helps exporters) / Neutral elsewhere
No other public model does this.
Key Numbers
| Languages | 10 (Arabic, German, English, Spanish, French, Hindi, Japanese, Korean, Portuguese, Chinese) |
| Training data | 149,481 perspective-labeled financial headlines |
| Base model | XLM-RoBERTa-large (560M parameters) |
| Labels | Negative / Neutral / Positive |
| Accuracy | 84.11% |
| F1 (macro) | 84.20% |
Quick Start
from transformers import pipeline
classifier = pipeline("text-classification", model="Kenpache/flame2")
# English — US investor perspective
classifier("[EN] Apple reported record quarterly revenue of $124 billion")
# [{'label': 'positive', 'score': 0.96}]
# Arabic — Gulf investor perspective
classifier("[AR] أسعار النفط تنخفض إلى 65 دولارا للبرميل")
# [{'label': 'negative', 'score': 0.93}] (oil down = bad for exporters)
# Hindi — Indian investor perspective
classifier("[HI] तेल की कीमतें गिरकर 65 डॉलर प्रति बैरल हुईं")
# [{'label': 'positive', 'score': 0.91}] (oil down = good for importers)
# Japanese
classifier("[JA] 日経平均株価が大幅下落、米中貿易摩擦の懸念で")
# [{'label': 'negative', 'score': 0.94}]
# Korean
classifier("[KO] 삼성전자 실적 호조에 코스피 상승")
# [{'label': 'positive', 'score': 0.92}]
# Chinese
classifier("[ZH] 中国央行降息50个基点，股市应声上涨")
# [{'label': 'positive', 'score': 0.95}]
# German
classifier("[DE] DAX erreicht neues Allzeithoch dank starker Bankenergebnisse")
# [{'label': 'positive', 'score': 0.93}]
# French
classifier("[FR] La Bourse de Paris chute de 3% après les tensions commerciales")
# [{'label': 'negative', 'score': 0.91}]
# Spanish
classifier("[ES] El beneficio neto de la compañía creció un 25% interanual")
# [{'label': 'positive', 'score': 0.94}]
# Portuguese
classifier("[PT] Ibovespa fecha em alta com otimismo sobre reforma tributária")
# [{'label': 'positive', 'score': 0.90}]
Important: Always use the [LANG] prefix ([EN], [AR], [HI], [JA], etc.) — this tells the model which market perspective to apply.
Supported Languages & Training Data
| Language | Code | Primary Economy | Oil Role | Total | Negative | Neutral | Positive |
|---|---|---|---|---|---|---|---|
| Arabic | AR | Gulf States (Saudi, UAE) | Exporter | 14,481 | 2,812 (19.4%) | 6,156 (42.5%) | 5,513 (38.1%) |
| German | DE | Germany / Eurozone | Importer | 15,000 | 3,544 (23.6%) | 6,636 (44.2%) | 4,820 (32.1%) |
| English | EN | United States | Mixed | 15,000 | 3,088 (20.6%) | 7,649 (51.0%) | 4,263 (28.4%) |
| Spanish | ES | Spain / Latin America | Importer | 15,000 | 3,872 (25.8%) | 5,616 (37.4%) | 5,512 (36.7%) |
| French | FR | France / Eurozone | Importer | 15,000 | 3,218 (21.5%) | 6,252 (41.7%) | 4,530 (30.2%) |
| Hindi | HI | India | Importer | 15,000 | 3,543 (23.6%) | 5,902 (39.3%) | 5,555 (37.0%) |
| Japanese | JA | Japan | Importer | 15,000 | 3,472 (23.1%) | 5,897 (39.3%) | 5,631 (37.5%) |
| Korean | KO | South Korea | Importer | 15,000 | 3,290 (21.9%) | 6,648 (44.3%) | 5,062 (33.7%) |
| Portuguese | PT | Brazil / Portugal | Exporter | 15,000 | 3,170 (21.1%) | 7,463 (49.8%) | 4,367 (29.1%) |
| Chinese | ZH | China | Importer | 15,000 | 3,542 (23.6%) | 4,055 (27.0%) | 7,403 (49.4%) |
Total: 149,481 labeled headlines across 10 languages.
Overall Class Distribution
| Class | Samples | Share |
|---|---|---|
| Negative | 33,551 | 22.4% |
| Neutral | 62,274 | 41.7% |
| Positive | 52,656 | 35.2% |
Data sources include financial news sites, stock market reports, and economic news agencies — labeled with perspective-aware rules specific to each economy.
What Makes FLAME2 Different
The Problem
Existing financial sentiment models treat sentiment as universal. But financial sentiment is not universal — it depends on where you are:
- Oil prices drop? Bad for Saudi Arabia, great for India.
- Yen weakens? Good for Japanese exporters, bad for Korean competitors.
- Fed raises rates? Bad for US stocks, often neutral for European markets.
Our Solution: Perspective-Aware Labels
Every headline in our dataset was labeled from the perspective of a local investor in that language's primary economy. The model learns that [AR] means "Gulf investor" and [HI] means "Indian investor."
Oil Price Rules
| Market Type | Oil Price Falls | Oil Price Rises | OPEC+ Output Increase |
|---|---|---|---|
| Exporters (AR, PT) | Negative | Positive | Negative |
| Importers (HI, KO, DE, FR, ES, JA, ZH) | Positive | Negative | Positive |
| Mixed (EN/US) | Positive | Context-dependent | Positive |
Currency Rules
| Language | Local Currency Strengthens | Local Currency Weakens |
|---|---|---|
| AR, PT, HI, KO, ZH | Positive | Negative |
| JA (export-driven) | Negative (hurts exporters) | Positive (helps exporters) |
| EN, DE, FR, ES | Neutral | Neutral |
Central Bank Rules
- Home central bank: rate cut = Positive, rate hike = Negative, hold = Neutral
- Foreign central bank: Neutral (unless headline explicitly links to local market impact)
Labels
| Label | ID | Examples |
|---|---|---|
| negative | 0 | Stock decline, losses, layoffs, downgrades, sanctions, bankruptcy |
| neutral | 1 | Factual reporting, mixed signals, foreign data without local impact |
| positive | 2 | Revenue growth, market rally, upgrades, new launches, rate cuts |
Results
Overall
| Metric | Score |
|---|---|
| Accuracy | 84.11% |
| F1 (macro) | 84.20% |
Per-Language Performance
| Language | Code | Accuracy | F1 Macro | Test Samples |
|---|---|---|---|---|
| Hindi | HI | 89.33% | 89.15% | 1,125 |
| Spanish | ES | 85.44% | 85.31% | 1,573 |
| Japanese | JA | 84.42% | 84.23% | 1,489 |
| French | FR | 84.06% | 84.24% | 2,579 |
| English | EN | 83.84% | 83.74% | 1,875 |
| Korean | KO | 83.54% | 83.71% | 3,280 |
| German | DE | 83.56% | 83.96% | 1,928 |
| Chinese | ZH | 83.50% | 81.43% | 1,751 |
| Portuguese | PT | 83.28% | 82.95% | 1,639 |
| Arabic | AR | 83.18% | 83.26% | 2,569 |
Per-Class Performance
| Class | Precision | Recall | F1 | Support |
|---|---|---|---|---|
| Negative | 0.81 | 0.87 | 0.84 | 4,487 |
| Neutral | 0.86 | 0.78 | 0.82 | 8,398 |
| Positive | 0.84 | 0.90 | 0.87 | 6,923 |
Training Pipeline
FLAME2 was built in two stages:
Stage 1: Supervised Fine-Tuning
XLM-RoBERTa-large was fine-tuned on ~150,000 perspective-labeled headlines with:
- Focal Loss (gamma=2.0) — focuses training on hard, misclassified examples instead of easy ones
- Class weights to handle label imbalance across languages
- Label smoothing (0.1) to handle ~3-5% annotation noise
- Language prefix
[LANG]injected before each headline for perspective routing - GroupShuffleSplit by news source domain — no article from the same source appears in both train and test (prevents data leakage)
- Gradient clipping (max_norm=1.0) for training stability
Stage 2: Live Stochastic Weight Averaging (SWA)
After epoch 12, the learning rate switches to a constant low rate (1e-5) and an AveragedModel maintains a running average of weights updated every epoch. This produces smoother, more generalizable predictions than any single checkpoint.
Training Details
| Parameter | Value |
|---|---|
| Base model | xlm-roberta-large (560M params) |
| Fine-tuning data | ~150,000 labeled headlines |
| Languages | 10 |
| Loss function | Focal Loss (gamma=2.0) |
| Learning rate | 2e-5 (→ 1e-5 SWA phase) |
| Label smoothing | 0.1 |
| Batch size | 32 |
| Max sequence length | 128 tokens |
| Precision | FP16 (mixed precision) |
| Train/Val/Test split | 70% / 15% / 15% |
| Split strategy | GroupShuffleSplit by source domain |
| SWA | Live averaging from epoch 12 |
Batch Processing
from transformers import pipeline
classifier = pipeline("text-classification", model="Kenpache/flame2", device=0)
texts = [
"[EN] Stocks rallied after the Fed signaled a pause in rate hikes.",
"[EN] The company filed for Chapter 11 bankruptcy protection.",
"[DE] DAX erreicht neues Allzeithoch dank starker Bankenergebnisse",
"[FR] La Bourse de Paris chute de 3% après les tensions commerciales",
"[ES] El beneficio neto de la compañía creció un 25% interanual",
"[ZH] 中国央行降息50个基点，股市应声上涨",
"[PT] Ibovespa fecha em alta com otimismo sobre reforma tributária",
"[AR] ارتفاع مؤشر السوق السعودي بنسبة 2% بعد إعلان أرباح أرامكو",
"[HI] भारतीय रिजर्व बैंक ने रेपो रेट में 25 बीपीएस की कटौती की",
"[JA] トヨタ自動車の純利益が前年比30%増加",
"[KO] 삼성전자 실적 호조에 코스피 상승",
]
results = classifier(texts, batch_size=32)
for text, result in zip(texts, results):
print(f"{result['label']:>8} ({result['score']:.2f}) {text[:70]}")
Use Cases
- Global News Monitoring — real-time sentiment classification across 10 markets
- Algorithmic Trading — perspective-aware signals: same event, different trades per market
- Portfolio Risk Management — track sentiment shifts across international holdings
- Cross-Market Arbitrage — detect when markets react differently to the same news
- Financial NLP Research — first multilingual perspective-aware sentiment benchmark
Limitations
- Optimized for news headlines (short text, 1-2 sentences). May underperform on long articles or social media.
- Perspective rules cover major economic patterns (oil, currency, central banks). Niche sector-specific effects may not be captured.
- Labels reflect the perspective of the primary economy for each language (e.g., AR = Gulf States, not all Arabic-speaking countries).
Citation
@misc{flame2_2026,
title={FLAME2: Financial Language Analysis for Multilingual Economics v2},
author={Kenpache},
year={2026},
url={https://huggingface.co/Kenpache/flame2}
}
License
Apache 2.0
- Downloads last month
- 65