Nasdaq Pvt Ltd is a global financial technology company that provides trading, analytics, and market infrastructure solutions. To enhance AI financial modeling capabilities, GoML developed a Physics Informed Neural Network based option pricing system by migrating a legacy MATLAB model to PyTorch and deploying it on AWS.
Problem: legacy systems limit AI financial modeling performance
Traditional option pricing systems relied on MATLAB based models, which limited scalability and slowed down AI financial modeling workflows. The system struggled to efficiently process large datasets, had high computation time due to CPU based training, and lacked support for multi GPU or distributed workloads.
It was also difficult to integrate with modern cloud infrastructure, and there was no capability for real time inference. These limitations reduced overall model performance and made it challenging to scale AI financial modeling effectively in production environments.
Solution: scalable AI financial modeling using PINN and cloud infrastructure
GoML built a cloud native ai financial modeling system using a Physics Informed Neural Network, powered by an AI Data Analytics Accelerator. The model embeds financial equations directly into training, improving prediction accuracy and reliability.
The system uses PyTorch for flexibility and AWS SageMaker for scalable training and deployment. It supports both CPU and GPU training along with distributed learning, enabling efficient large scale data processing and advanced AI financial modeling workflows.
Model architecture for AI financial modeling
The core of the solution is a PINN based neural network designed for option pricing.
Key improvements
- Handles financial PDE constraints directly during training
- Transformer inspired architecture with dual encoders
- 12 hidden layers with high dimensional capacity
- Around 25 million trainable parameters
- Improved accuracy through physics guided learning
Training pipeline and scalability
The solution introduces a flexible and scalable training pipeline.
Key capabilities
- Supports CPU, single GPU, and multi GPU training
- Multi node distributed training using PyTorch DDP
- Automated job orchestration using SageMaker
- Checkpointing and resume support for long runs
- Spot instances for cost optimized training
Inference and deployment
The model is deployed using a serverless inference pipeline for real time predictions.
- Key capabilities
- Serverless endpoints with auto scaling
- Low latency prediction APIs
- Support for both batch and real time inference
- Flexible deployment from local or S3 model artifacts
Cloud infrastructure and system design
The solution uses a secure and scalable AWS architecture.
Key components
- SageMaker for training and inference orchestration
- S3 for storing datasets and model artifacts
- ECR for container management
- CloudWatch for monitoring and logging
- VPC and endpoints for secure networking
Operational efficiency and automation
The system improves developer productivity and operational efficiency.
Key improvements
- Centralized configuration using YAML
- Automated training job submission
- Real time monitoring of training and inference
- Fault tolerance with checkpoint recovery
- Simplified deployment workflows
.svg)
Impacts
- 12x lower MAE compared to legacy model
- 8x lower MAPE
- 3x lower sMAPE
- R squared improved to near perfect accuracy
- 5x faster training
- 70% cost savings with optimized infrastructure
About
Before Gen AI and after Gen AI
“With AI financial modeling powered by PINN, Nasdaq improves accuracy, scalability, and speed in option pricing, enabling better financial decision making.”
Prashanna Rao, Head of Engineering, GoML
Key takeaways for AI financial modeling
Common challenges
- Legacy systems do not scale
- High computation cost
- Lack of real time prediction systems
- Difficulty handling complex financial equations
Practical guidance
- Adopt ai financial modeling with neural networks
- Use physics informed models for better accuracy
- Leverage GPU training for faster results
- Deploy serverless inference for cost efficiency
- Build cloud native infrastructure for scalability
Ready to build ai financial modeling solutions
Partner with GoML to build scalable AI financial modeling systems using PINN, deep learning, and cloud infrastructure with AI Matic.
