Fin-Edge: Financial Sentiment Analysis on Edge Devices

Fin-Edge: Compressing FinBERT for real-time financial sentiment on edge devices with 10× faster inference and ~90% smaller models while preserving accuracy.

Comment

Inspiration

  • This project started from the tension between powerful transformer models like FinBERT and the practical need to run them on edge hardware where latency, cost, and energy matter. ​
  • The hackathon’s focus on real-world impact pushed the idea of “FinBERT-in-your-pocket” for mobile traders, RM apps, and on-device risk alerts. ## What it does
  • Fin-Edge runs a two-stage pipeline that first trains a baseline FinBERT sentiment classifier on the Financial PhraseBank and then generates a compressed “Fin-Edge” model tailored for edge devices.
  • It measures and compares model size, single-sentence latency, and accuracy, achieving a projected drop from about 417.7 MB to 48.0 MB and 4.63 ms to 0.56 ms while keeping accuracy around the low 80% range. ## How we built it -Implemented a full training/evaluation loop in PyTorch using the Financial PhraseBank “50% agreement” split, with a custom FinancialDataset class and DataLoaders for GPU-backed training. -​Integrated perforatedai with HuggingFace’s BertForSequenceClassification, then instrumented model size, single-sentence inference latency, and accuracy to log metrics and export charts.

Challenges we ran into

  • PerforatedAI’s dendritic optimization surfaced multiple this_output_dimensions shape mismatches across deeper BERT layers, forcing a fallback to a projected performance mode instead of a fully trained compressed model. ​- Balancing hackathon time limits, GPU quota, brittle beta-library behavior, and still producing clean charts, JSON metrics, and a reproducible notebook pipeline required disciplined experimentation and careful error handling.

Accomplishments that we're proud of

  • Delivered a full FinBERT-to-edge pipeline that trains a baseline model and generates a compact “Fin-Edge” variant with deployment-ready charts and JSON metrics.
  • Achieved a projected reduction from about 417.7 MB to 48.0 MB and latency from 4.63 ms to 0.56 ms while keeping accuracy in the low 80% range.

What we learned

  • Edge deployment is not just about compressing parameters; it is about balancing model size, latency, and accuracy under strict hardware constraints.
  • Working with a beta optimization library (PerforatedAI) taught how to design robust fallbacks and still produce meaningful “industry-scale” projections when full compression is not stable.

What's next for Fin-Edge: Financial Sentiment Analysis on Edge Devices

  • Move from projected performance to a fully stable compressed FinBERT by resolving dendritic shape-mismatch issues layer by layer.
  • Package Fin-Edge as an on-device SDK for financial apps, with configurable latency–accuracy trade-offs and support for more financial sentiment datasets.

Built With

Share this project:

Updates