Finance Forecasting LLM Development
Budget: €30 – €250 EUR
I need a purpose-built large language model that performs end-to-end financial forecasting. The model must comfortably switch between stock-market prediction, company-level revenue projections, and broader economic trend analysis without retraining or manual prompt engineering.
Data the model must ingest and reason over:
• Historical market data
• Company financial reports
• Economic indicators
I will supply sample datasets as CSV, XBRL, and API endpoints; you may suggest additional publicly available feeds if they strengthen performance.
Scope
• Design the data pipeline, including cleaning, feature extraction, and secure storage.
• Fine-tune or train an LLM (e.g., GPT-J, Llama-2, or a comparable open-source model) to handle multimodal numeric-text inputs.
• Implement evaluation routines—back-testing for market forecasts, MAPE or SMAPE for revenue projections, and directional accuracy on macro indicators.
• Expose the model through a lightweight REST or gRPC service with clear inference examples in Python.
• Provide concise documentation covering setup, data refresh, and prompt patterns.
Acceptance Criteria
The delivered model must:
• Demonstrate statistically significant predictive power versus naïve baselines on held-out data.
• Run reproducibly on a single GPU (A100 or lower) or an M1-class CPU with quantization.
• Return responses within two seconds for typical prompts.
Tools such as PyTorch, TensorFlow, HuggingFace Transformers, LangChain, or equivalent are perfectly fine; please choose what lets you iterate fastest while keeping the stack transparent.
Timeline and collaboration style are flexible, though I would like a short kickoff chat and weekly progress checkpoints. Let me know which model architecture you have in mind and any comparable work you can share.
---
Ideas are welcome! I just need an LLM in finance.
Data the model must ingest and reason over:
• Historical market data
• Company financial reports
• Economic indicators
I will supply sample datasets as CSV, XBRL, and API endpoints; you may suggest additional publicly available feeds if they strengthen performance.
Scope
• Design the data pipeline, including cleaning, feature extraction, and secure storage.
• Fine-tune or train an LLM (e.g., GPT-J, Llama-2, or a comparable open-source model) to handle multimodal numeric-text inputs.
• Implement evaluation routines—back-testing for market forecasts, MAPE or SMAPE for revenue projections, and directional accuracy on macro indicators.
• Expose the model through a lightweight REST or gRPC service with clear inference examples in Python.
• Provide concise documentation covering setup, data refresh, and prompt patterns.
Acceptance Criteria
The delivered model must:
• Demonstrate statistically significant predictive power versus naïve baselines on held-out data.
• Run reproducibly on a single GPU (A100 or lower) or an M1-class CPU with quantization.
• Return responses within two seconds for typical prompts.
Tools such as PyTorch, TensorFlow, HuggingFace Transformers, LangChain, or equivalent are perfectly fine; please choose what lets you iterate fastest while keeping the stack transparent.
Timeline and collaboration style are flexible, though I would like a short kickoff chat and weekly progress checkpoints. Let me know which model architecture you have in mind and any comparable work you can share.
---
Ideas are welcome! I just need an LLM in finance.