FinBERT Trading
Serverless NLP and trading data platform with sentiment analysis as decision support.
An engineering case for Applied AI / Financial Data: automated data pipelines, German FinBERT for sentiment analysis and reinforcement-learning-based paper trading – as decision support, not as a promise of profitable automated trading.
- Technical challenge
- NLP models and financial data were to be combined into robust decision support – as an end-to-end data and ML pipeline, with an honest assessment of the limits rather than exaggerated promises.
- Architecture
- A fully serverless AWS platform (CDK): RSS pipeline and German FinBERT for sentiment analysis, data preparation with training including walk-forward validation, and automated paper trading with risk management on 16-feature observations.
- Engineering challenges
- Data quality and consistency across many sources, traceable model evaluation (walk-forward instead of backtest bias) and stable 24/7 operation without manual provisioning.
- Demonstrated capabilities
- NLP and sentiment analysis, reinforcement-learning pipelines, automated data engineering and serverless AWS architecture with CDK.
FinBERT Trading is a Riegel Systems engineering case in the field of Applied AI / Financial Data. It demonstrates how NLP models, financial data and automated pipelines can be combined into robust decision support – as a fully serverless system on AWS.
Structure
- Data sources (RSS + ticker)
- FinBERT sentiment analysis
- Data preparation & features
- PPO training (walk-forward)
- Paper trading with risk management
- Dashboard
Architecture
Engineering focus
- Sentiment pipeline: A German FinBERT model continuously classifies news from multiple sources; results are cached with TTL.
- Training with walk-forward validation: Instead of naive backtesting, model quality is evaluated honestly over time – a core element for reliable statements.
- Paper Trading: The agent trades in simulated 5-minute intervals with 16-feature observations, five actions and explicit risk management – without real capital.
- Serverless operation: Infrastructure as code (CDK), scale-to-zero, no manual provisioning.
Honest assessment
Financial markets cannot be reliably predicted by data alone. The value of this case lies in the robust engineering of the data and analysis paths as well as in the traceable evaluation of the model’s limits – explicitly without any investment or return promises.
Technologies & concepts
A similar engineering project?
If your project needs comparable technical depth, we discuss feasibility, architecture and effort – concretely rather than in pitch format.