Machine Learning

Machine learning inside your quant library

Put a neural network inside your C++ analytics and differentiate through both. For models up to about 1,000 inputs AADC trains and runs faster on CPU than TensorFlow, and it deploys without Python.

Schedule Demo See Benchmarks

The problem

Python ML frameworks are slow in a production pricing path, and rewriting a trained model in C++ by hand is slow in a different way. AADC lets the model and the analytics around it be one compiled kernel.

Where the big frameworks are a poor fit

  • CPU Performance: TensorFlow/PyTorch optimised for GPUs, not production CPU systems
  • Integration Issues: Difficult to integrate with existing C++ analytics
  • Model Size Mismatch: Over-engineered for moderate-sized quant models (up to 1,000 inputs)
  • Deployment Complexity: Python dependencies in production environments
  • Training Overhead: Slower training cycles for iterative model development
  • No Business Logic Integration: Hard to mix a network with the domain calculation around it

AADC vs TensorFlow: CPU Performance Comparison

ADBench benchmark results show AADC's significant advantage for quantitative finance applications

ADBench Performance Comparison: AADC vs TensorFlow vs PyTorch

ADBench Performance Comparison: AADC vs TensorFlow vs PyTorch

Key Performance Benefits:

10x+
Faster than TensorFlow
On CPU for typical quant models
3x
More Accurate
Time series forecasting, as reported in the Olsson paper
Several x
Faster Training
For neural network development

Note: For very large models, in vision or language, performance is bound by memory bandwidth and the Python frameworks are the right tool. The quant range, hundreds to a thousand inputs, is where AADC is faster.

Research & Publications

State-of-the-art neural network research from MatLogica and partners

ARN: Automatic Synthesis of Neurons for Recurrent Neural Nets

Novel ARN neurons achieve up to 3x better accuracy than LSTMs on time series tasks through automatic synthesis of recurrent architectures with skip connections and novel activation functions.

View Paper

Automatic Synthesis of Recurrent Neurons for Imitation Learning From CNC Machine Operators

IEEE paper exploring innovative automatic synthesis of recurrent neurons to enable imitation learning directly from expert CNC machine operators.

View Paper

Recurrent Neural Networks for Oil Well Event Prediction

IEEE Intelligent Systems publication evaluating recurrent neural network architectures for predicting anomalies and events in oil well operations using the 3W benchmark dataset.

View Paper

Integration Approaches

Choose the approach that fits your development workflow

Approach 1

Standalone ML Models

Build neural networks in C++ with AADC for pure ML workloads

  • Replace TensorFlow/PyTorch for CPU deployment
  • 10x+ faster training and inference
  • No Python runtime dependencies
  • Standard neural network layers available
Approach 2

Hybrid Analytics + ML

Interleave a network with the business logic around it

  • ML layers embedded in pricing/risk models
  • Single differentiation pass through entire pipeline
  • Domain calculations + ML predictions unified
  • Natural C++ code integration
Approach 3

Differential Machine Learning

Differential regression and differential PCA train on the pathwise differentials of the pricing function.

  • Differential labels from the same recording as the price
  • Every sample carries its full gradient
  • Trains on C++ or Python pricing code as written
  • No separate differentiation code to maintain

Why AADC Excels at ML for Quantitative Finance

Optimised for Model Size

TensorFlow and PyTorch are designed for massive models (millions of parameters). Quantitative finance typically needs models with hundreds to thousands of.

CPU-First Architecture

Most ML frameworks optimise for GPUs. AADC is built for CPU performance, which is where production quant systems run. No GPU infrastructure needed.

Zero Framework Overhead

No graph building overhead, no Python interpreter, no framework abstractions. Direct compilation to optimised machine code with automatic differentiation.

Natural C++ Integration

ML layers sit inside your existing C++ analytics. One language, no serialisation between the two, one binary to deploy.

AADC vs Traditional ML Frameworks

For quantitative finance applications

Feature TensorFlow/PyTorch AADC Advantage
CPU Performance (up to 1K inputs) Baseline 10x+ faster AADC
Training Speed Baseline Several times faster AADC
Integration with C++ Analytics Complex (language barriers) Native C++ AADC
Production Deployment Python runtime required Native C++ binary AADC
GPU Requirement Needed for good performance CPU-only AADC
Custom Layer Development Framework-specific APIs Standard C++ code AADC
Very Large Models (>10K inputs) Better Good TensorFlow/PyTorch
Pre-trained Model Ecosystem Extensive Limited TensorFlow/PyTorch

Note: AADC excels in the quantitative finance sweet spot (moderate model sizes, CPU deployment, integration with business logic), while TensorFlow/PyTorch are better for very large models and transfer learning scenarios.

Ready to Accelerate Your ML Workflows?

See how AADC can deliver 10x+ performance improvements for your machine learning applications in quantitative finance.

Schedule Consultation Contact Us