š¬ FX Market Edge Prediction ā¢ Machine Learning Research
š§ Project Overview
Comprehensive investigation into USD pip movement predictability using hybrid Temporal Convolutional Autoencoders (TCNAE)
with LightGBM across 3 years of hourly data from 24 major currency pairs. Despite sophisticated methodology,
no statistically significant market edge was discovered, providing valuable negative evidence supporting
the efficient market hypothesis.
šļø Technical Architecture
- TCNAE Model: 537K-parameter temporal convolutional autoencoder compressing 4-hour sequences into 120-dimensional representations
- LightGBM Pipeline: 48 specialized gradient boosting models (2 per instrument) for pip magnitude and direction prediction
- Cross-Instrument Context: 24Ć5 feature tensor enabling information sharing across currency pairs
- Production Integration: OANDA v20 API, Docker containerization, real-time prediction pipeline
š Dataset & Results
- Scale: 224,955 samples across 24 currency pairs over 3 years
- Performance: 50.1% direction accuracy (random baseline), 0.02% average correlation
- Methodology: Rigorous temporal validation, aggressive quality filtering (50% retention)
- Innovation: Dual model architecture, latent caching system (70% computational overhead reduction)
šÆ Scientific Contribution
This study provides crucial negative evidence using modern ML techniques, serving as a methodological
template for rigorous financial prediction research. The honest reporting of unsuccessful edge discovery attempts
contributes valuable knowledge validating market efficiency at hourly timeframes.
š ļø Technology Stack
Python
PyTorch
LightGBM
OANDA v20 API
Docker
Financial ML
Repository Features: Production-ready code, comprehensive documentation, experimental results,
trained models, visualization suite, and Docker deployment configuration.
Project Details: View Detailed Project Page
Code Repository: FX Market Edge Prediction Research
š MuZero Trading Pipeline ā¢ JanāJul 2025
š§ Overview
Built a productionāready MuZeroābased trading system capable of processing 7,876āÆticks/sec, calculating 264 technical indicators, and delivering <100āÆms latency via a realātime dual dashboard.
ā± Development Timeline
- JanāFeb: Core MuZero prototype, Kafka ingestion, basic indicator calculations.
- MarāApr: Episode collection, reward logic, trade execution implemented.
- May: Terminal & WebSocket dashboards added; expanded features; production hardening.
- Jun: LightZero C++ MCTS integrated; streaming migrated to Redpanda & SQLite; speed optimized.
- Jul: Released V5 with singleācommand deployment; removed pandas/pickle; optimized pipeline with IoTDB & SQLiteāLZ4.
š Measured Outcomes
- Throughput: 245 ā 7,876āÆticks/sec (10Ć gain)
- Memory: 8.2āÆGB ā 3.2āÆGB (60% reduction)
- Training time: 3Ć faster
- Monitoring latency: <100āÆms via WebSocket
š§ Core Contributions
- Engineered a pipeline with 264 multiātimeframe indicators (tickāD1).
- Designed V5 model: lightweight MLP with optimized C++ MCTS integration.
- Built robust instrumentation: dual dashboards, SQLite logging, autoāresume, cleanup.
- Eliminated technical debtāstreamlined all components to production standard.
šÆ Development Principles
- Prioritized simplicity and clarity; avoided overāengineering.
- Followed measureābeforeāoptimize approach; profiling guided improvements.
- Maintained a productionāfirst mindset; built with reliability and observability in focus.
This project reflects a complete journey from prototype to production, showcasing performance gains, reliable monitoring, and a clean codebaseāan example of disciplined, solo engineering at scale.
OANDA NEAT Trading Algorithm
Role: Self-Directed Project
Description: Developed a trading algorithm that utilizes NeuroEvolution of Augmenting Topologies (NEAT) for decision-making in the forex market on the OANDA trading platform. The algorithm is designed to learn from historical market data to evolve and optimize trading strategies through genetic algorithms.
The genetic algorithm dynamically evolves neural network architectures to find solutions, increasing in complexity only as needed to reach goal behaviors. A carefully designed reward function encourages profitable trading behaviors while managing risk and optimizing trade frequency.
Initial results include performance visualizations and trade logs, showcasing the algorithm's capabilities and areas for further exploration.
Technologies Used: Python, genetic algorithms, neural networks, data analysis.
Code Repository: OANDA NEAT Trading Algorithm
Peoplesā FinTech Expectancy Calculator
Role: Founder and CTO, Peoples' FinTech
Description: I led the development of the Peoplesā FinTech Expectancy Calculator, a practical tool designed to help traders evaluate the long-term profitability of their trading strategies. The calculator provides insights into the expected return for every dollar risked and offers a visual representation of moving window expectancy to help users assess the stability of their performance. You can explore the tool here. Additionally, thereās a detailed article titled The Critical Importance of Expectancy as a Strategy Performance Metric, which you can read here.
Key Contributions:
- Product Vision: Created the Expectancy Calculator to address a significant need in trading strategy assessment, focusing on helping traders understand expectancy as a key performance metric based on Dr. Van Tharpās framework.
- Technical Architecture: Oversaw the technical design of the tool, ensuring user-friendly input methods and dynamic data visualizations that enhance the user experience.
- Strategic Impact: Positioned the Expectancy Calculator as a valuable resource for trader education, aiding users in recognizing common psychological and statistical biases through clear, quantitative insights into their strategy's potential.
Skills: Strategic product development, financial metrics, JavaScript, web-based tool design, trading system evaluation, team leadership.
Habirokrator Project
Role: Web Developer (Client Project)
Description: Collaborated with the vision holder to design and develop a user-friendly website offering consulting services for navigating Israel's traffic violation judiciary and bureaucracy. The website provides resources such as guidelines for disputing traffic tickets, parking fines, toll road information, and speed camera locations.
Implemented intuitive navigation and functional design to ensure accessibility across devices and browsers. Enhanced user experience by integrating WhatsApp for seamless communication with the consulting service.
Technologies Used: HTML, CSS, JavaScript, W3.CSS framework, WhatsApp integration.
Project Link: Habirokrator
Anitrade Project
Role: Self-Directed at Anitani Solutions LLC
Description: Conducted research and development of algorithmic trading systems, resulting in a collection of technical indicators, trading robots, and expert advisors for the MetaTrader platform, along with C# projects for automated trading on cTrader.
Developed a proof-of-concept using feedforward neural networks to predict minimal daily moves in the forex market. This involved C# code integrating a Python script for predictions using PyBrain. Code Here
Created a live trading robot for a range-breakout strategy in the spot forex market, achieving a 50% return over two and a half months during two trial periods (annualized return of 619%).
Technologies Used: FIX, machine learning, time series analysis, TradeStation, MetaTrader, cTrader/cAlgo.
Code Repository: Anitrade Contributed Codebase