NSE LLaMA Sentiment Forecast Pipeline
Budget: ₹37,500 – ₹75,000 INR
I want a production-ready pipeline that ingests official NSE reports along with any additional unstructured market text I provide, then uses a deep learning-based sentiment model—fine-tuned from LLaMA—to score each document, engineer features, and predict equity returns over six horizons (1, 3, 7, 15, 30 and 60 days).
Here’s the flow I have in mind: raw PDFs or HTML from the exchange land in cloud storage; they are parsed, cleaned and chunked, then passed through a LLaMA sentiment head that outputs a numeric score plus an explanation token stream. Those scores join technical and fundamental factors in a feature store, after which a multi-horizon regression network trains nightly and pushes updated forecasts to an API endpoint and dashboard.
What I need from you:
• Code (Python) for every stage—ingestion, preprocessing, LLaMA fine-tuning, vectorisation, feature engineering and model training/evaluation.
• A concise inference service (FastAPI or similar) that returns sentiment, engineered features and the nine-step return predictions.
• Clear documentation and a one-click deployment script (Docker/Compose or Terraform) so I can spin the whole stack up on my GPU instance.
Use whichever deep-learning framework you are most comfortable with; PyTorch or TensorFlow are both fine. Accuracy on a hold-out set and reproducible results will be the acceptance criteria. If this first phase runs smoothly, there is follow-on work to expand coverage and integrate streaming social data as well.
Here’s the flow I have in mind: raw PDFs or HTML from the exchange land in cloud storage; they are parsed, cleaned and chunked, then passed through a LLaMA sentiment head that outputs a numeric score plus an explanation token stream. Those scores join technical and fundamental factors in a feature store, after which a multi-horizon regression network trains nightly and pushes updated forecasts to an API endpoint and dashboard.
What I need from you:
• Code (Python) for every stage—ingestion, preprocessing, LLaMA fine-tuning, vectorisation, feature engineering and model training/evaluation.
• A concise inference service (FastAPI or similar) that returns sentiment, engineered features and the nine-step return predictions.
• Clear documentation and a one-click deployment script (Docker/Compose or Terraform) so I can spin the whole stack up on my GPU instance.
Use whichever deep-learning framework you are most comfortable with; PyTorch or TensorFlow are both fine. Accuracy on a hold-out set and reproducible results will be the acceptance criteria. If this first phase runs smoothly, there is follow-on work to expand coverage and integrate streaming social data as well.