ATLAS: Bringing Medical AI Innovation to Financial Markets

For my ICS4U Computer Science final project, I developed ATLAS (Advanced Technical Learning Analysis System), an AI-powered stock market pattern recognition system that achieves 83.4% validation accuracy with remarkable parameter efficiency.

The Core Innovation

ATLAS makes a bold cross-domain leap: applying nnU-Net's revolutionary auto-configuration approach from medical image segmentation to financial time series prediction. Just as nnU-Net transformed medical AI by eliminating manual architecture design, ATLAS eliminates the expertise barrier in financial machine learning through intelligent automation.

What Makes ATLAS Different

Knowledge Distillation from Human Expertise

Traditional CNNs struggle to learn financial chart patterns effectively. ATLAS solves this by directly encoding trading expertise into specialized convolution kernels—essentially teaching the neural network to recognize head-and-shoulders patterns, support/resistance levels, and breakout signals the way experienced traders do.

nnU-Net-Inspired Auto-Tuning

The system automatically analyzes market characteristics (volatility, trend strength, noise levels) and intelligently configures hyperparameters without manual intervention. High volatility markets get larger context windows and conservative learning rates; stable markets get aggressive optimization for faster convergence.

This approach proved remarkably effective: 3.6-6.8% accuracy improvements across different market sectors with zero manual tuning.

Performance Highlights

Accuracy and Efficiency

83.4% validation accuracy with only 17,081 parameters
18.5× more efficient than ResNet50 (which has 316,641 parameters)
84.8% accuracy on minute-level data for high-frequency trading scenarios

Cross-Platform Performance

Sub-millisecond inference (0.94ms on H100 GPU)
Consistent accuracy across NVIDIA, Huawei Ascend, AMD, and Intel hardware
29,269 samples/second throughput on enterprise hardware

Robust Validation

Walk-forward validation: 86.3% accuracy (most realistic for trading)
Time series cross-validation: 83.6% accuracy
Validated across 200+ stocks and multiple market sectors

Technical Architecture

Multi-Modal Time Series to Image Pipeline

ATLAS transforms stock price data into four complementary image representations:

GASF (Gramian Angular Summation Field): Captures overall trends
GADF (Gramian Angular Difference Field): Detects directional changes
RP (Recurrence Plot): Identifies repetitive patterns
MTF (Markov Transition Field): Recognizes state transitions

Each representation feeds into a specialized CNN branch with hand-crafted kernels designed to detect specific financial patterns. The system then fuses these multi-modal features for final prediction.

Auto-Tuning System

The auto-tuning engine analyzes dataset fingerprints and applies financial domain rules:

Volatile markets → larger windows, conservative learning
Clean data → aggressive optimization, sensitive thresholds
Limited data → higher regularization, extended training

Real-World Applications

High-Frequency Trading: Minute-level signal generation with 84.8% accuracy
Algorithmic Trading: Daily position signals for systematic strategies
Cross-Market Trading: Validated on both US stocks and Chinese A-shares
Portfolio Optimization: Multi-timeframe analysis for risk management

What I Learned

Building ATLAS taught me far more than just implementing a neural network:

Cross-Domain Innovation: The best solutions often come from unexpected places. Applying medical AI techniques to finance required deep understanding of both domains.
Parameter Efficiency Matters: In production systems, a model that's 18× smaller with minimal accuracy trade-off is often more valuable than a marginally better but resource-heavy model.
Automation is Engineering: The auto-tuning system represents hundreds of hours of research into what makes financial models work—codified into algorithmic rules that anyone can use.
Validation is Critical: Walk-forward validation revealed the importance of realistic testing. Traditional holdout methods can be misleading for time series data.

Technical Stack

Deep Learning: PyTorch with custom CNN architectures
Data Processing: Pandas, NumPy, SciPy for financial data
Technical Analysis: 14 indicators including RSI, MACD, Bollinger Bands, Kalman filters
Visualization: Plotly Dash for real-time multi-stock dashboard
Deployment: Cross-platform optimization for production readiness

Dataset

The system trained on 4+ years of historical data covering 200+ US stocks across technology, finance, healthcare, and other sectors. I published the preprocessed dataset on Kaggle to help other researchers: US Stock Collect Data.

Looking Back

This project represents the culmination of my interests in AI and quantitative finance. More importantly, it demonstrates that high school students can tackle sophisticated problems at the intersection of multiple domains—combining medical AI research, financial market theory, and production-grade software engineering.

The success of ATLAS validates an important principle: when you encode domain expertise into algorithms and automate the tedious parts, you democratize access to sophisticated techniques. That's the kind of engineering I want to pursue in my career.