Property	Value
Base model	ProsusAI/finbert (BERT-base fine-tuned on Reuters/Bloomberg)
Task	3-class sentiment: negative / neutral / positive
Fine-tuning strategy	Frozen layers 0–9, trainable layers 10–11 + pooler + head
Training data	~40,600 samples (FinGPT financial sentiment + Twitter Financial News + ~70 synthetic geopolitical headlines)
Class balancing	Undersampling + weighted CrossEntropyLoss (neg=1.459, neu=1.060, pos=0.729)
Test accuracy	0.6393
Best val accuracy	0.6454

Performance

Class	Precision	Recall	F1
negative	0.73	0.86	0.79
neutral	0.52	0.75	0.62
positive	0.74	0.45	0.56
overall	0.67	0.64	0.63

Labels

ID	Label	Meaning
0	negative	Risk increasing — conflict, sanctions, disaster, supplier failure
1	neutral	Routine updates, mixed signals, uncertainty
2	positive	Risk decreasing — stability, trade agreements, recovery

Usage

python

from transformers import pipeline

classifier = pipeline(
    "text-classification",
    model="arunabhachanda/supplychain-finbert",
    return_all_scores=True,
)

result = classifier("Ceasefire in the region reopens key supply corridors")
# → [{'label': 'negative', 'score': 0.04},
#    {'label': 'neutral',  'score': 0.11},
#    {'label': 'positive', 'score': 0.85}]

# Polarity score used by SupplyGuard AI:
polarity = result[2]['score'] - result[0]['score']   # P(positive) - P(negative)
# → float in [-1.0, +1.0]  used as region_news_sentiment feature

Transfer Learning Architecture

ProsusAI/finbert (pre-trained on financial news corpus)
├── BERT Embeddings          [FROZEN]      ← vocabulary + positional encoding
├── Transformer Layer 0–9    [FROZEN]      ← general language + financial knowledge
├── Transformer Layer 10–11  [TRAINABLE]   ← adapted to supply-chain language
├── Pooler                   [TRAINABLE]   ← [CLS] token representation
└── Classifier Head (768→3)  [TRAINABLE]   ← new head for 3-class sentiment

Trainable parameters: 14,768,643 (13.5% of total)
Frozen parameters: 94,715,904 (86.5% of total)

Training Details

Optimizer: AdamW (lr=2e-5, weight_decay=0.01)
Scheduler: Linear warmup (10% steps) + linear decay
Epochs: 4
Batch size: 16
Gradient clipping: max_norm=1.0
Class weights: neg=1.459, neu=1.060, pos=0.729 (weighted CrossEntropyLoss)
Split: 80% train / 10% val / 10% test (stratified)

Built By

Arunabha Kumar Chanda — M.Sc. Business Intelligence & Data Science, ISM Munich
GitHub: arunabhachanda