Machine Learning & Quantitative Research

Sharifi designs reinforcement-learning and event-driven trading systems for nonstationary markets, focusing on representation learning for regime discovery, attention-based sequence modeling, uncertainty quantification, and drawdown-sensitive objectives. His validation stack uses walk-forward analysis, purged and embargoed cross-validation, Monte Carlo resampling, and explicit transaction-cost and capacity modeling to ensure claims are leak-free and operationally realistic.

In AI and quantitative research, Sharifi designs high-level frameworks for machine learning, reinforcement learning, and predictive modeling on nonstationary financial time series. His work centers on representation learning for regime discovery, attention-based sequence models and state-space formulations for long-horizon dependencies, and risk-aware objectives that optimize drawdown-sensitive utility rather than point accuracy. He studies uncertainty quantification through ensembling and stochastic inference, applies leak-free evaluation with walk-forward analysis and purged and embargoed cross-validation, and uses meta-labeling with the triple-barrier method to separate entry logic from outcome conditioning. Transaction costs, capacity, and constraints are treated as integral parts of the objective. The same mathematical focus that underpins his music informs this research, including spectral and multiresolution analysis, change-point detection, and morphology-driven segmentation, yielding systems that connect formal structure, pattern finding, and rigorous validation across music and markets.

Case Study:

QuantRL: Deep Reinforcement Learning for Trading

Github Repository

QuantRL, is a modular, research-grade reinforcement learning framework for algorithmic trading. QuantRL is built around the Proximal Policy Optimization (PPO) algorithm, enhanced by a custom self-attention neural network that allows its AI trading agent to recognize intricate market patterns and make data-driven decisions. The system is designed as an end-to-end pipeline: it ingests historical market data, engineers a rich set of features, and trains an agent within a realistic simulated environment. This custom OpenAI Gym-based environment models real market conditions, including transaction costs, slippage, and portfolio factors, providing a safe sandbox for the AI to learn trading strategies. This modular pipeline also ensures any strategy learned by the agent is rigorously evaluated: QuantRL includes extensive backtesting against baseline strategies and tracks over 30 performance metrics to analyze outcomes from multiple angles.

Extensive Feature Engineering: QuantRL incorporates 50+ engineered features from technical indicators (e.g. RSI, MACD, Bollinger Bands) to volatility regimes and intraday patterns, giving the agent a comprehensive view of market state.
Advanced RL Agent: It leverages a PPO-based reinforcement learning agent with a self-attention network architecture, complete with residual connections and layer normalization, enabling the model to capture complex temporal dependencies in financial data.
Robust Simulation Environment: The framework supports both discrete and continuous trading actions with various position-sizing strategies (fixed lots, Kelly criterion, volatility scaling), while simulating realistic factors like slippage, transaction costs, and drawdowns. The environment continuously tracks portfolio metrics such as profit-and-loss, Sharpe ratio, and maximum drawdown to provide immediate feedback on performance.
Comprehensive Backtesting: After training, strategies are tested against benchmark approaches (e.g. buy-and-hold or moving-average crossover) using walk-forward and Monte Carlo simulations. QuantRL records over 30 performance metrics and generates rich visualizations (equity curves, rolling Sharpe ratios, drawdown plots) to validate each strategy’s robustness.

Developed as an independent project, QuantRL showcases Amin’s ability to integrate modern AI techniques with finance. It demonstrates production-level engineering and creativity in quantitative modeling, providing a modular, extensible platform where new ideas (such as alternative reward functions, multi-asset strategies, or ensemble methods) can be prototyped and evaluated in a rigorous research setting.

Comprehensive Backtest Analysis
Benchmarks the RL agent against SMA crossover and Buy & Hold strategies.

Sample Backtest Results (on AAPL 5-min data, 5 years)

This framework is in active development and not fully optimized. The backtest below was performed using an early version of the agent trained on limited data and compute resources.

Cumulative Return: ~7.4%
Sharpe Ratio: ~1.67
Sortino Ratio: ~1.74
Calmar Ratio: ~1.64
Max Drawdown: ~7.1%
Win Rate: ~54%
Trades Executed: ~280
Average Position Duration: 30–45 minutes

Further improvements are expected with larger datasets, longer training schedules, and hyperparameter tuning.

Case Study:

AlphaFX: Multi-Timeframe Automated Trading System

Github Repository

Complementing his work in reinforcement learning, Amin also built AlphaFX, a production-grade, event-driven trading system for live multi-timeframe forex trading. AlphaFX is designed to operate across multiple horizons, for example, simultaneously analyzing 1-minute, 5-minute, and 15-minute charts, to capture both fine-grained price movements and broader market trends. Built around OANDA’s v20 REST API for real-time data streaming, it integrates high-frequency data processing with sophisticated signal generation logic: at its core is a Complex Event Processing (CEP) engine that fuses indicators and patterns across different time scales, ensuring that short-term signals align with longer-term market context before executing trades. This system also includes tools for offline analysis, monitoring dashboards, and performance reporting to facilitate end-to-end deployment and evaluation.

Real-Time Data & Indicators: AlphaFX continuously fetches live market data (e.g. 1m, 5m, 15m price feeds) asynchronously, maintaining a synchronized view of multiple timeframes. It computes a broad array of technical indicators, from classic measures like moving averages, MACD, and RSI to custom signals capturing candlestick patterns and volume trends, all adaptively tuned to current market conditions.
Adaptive Signal Engine: The CEP-based signal engine evaluates patterns across timeframes, applying Gaussian Mixture Models for market regime detection. This regime awareness allows AlphaFX to adjust its strategy rules on the fly, using context-specific thresholds and filters to screen out noisy signals and confirm high-probability trade opportunities.
Risk Management & Execution: Emphasizing capital preservation, AlphaFX employs research-backed risk management strategies. It uses Kelly Criterion position sizing (scaled by volatility and confidence) and sets dynamic stop-loss/take-profit levels based on Average True Range (ATR), so that risk is calibrated to market volatility. The system also enforces strict limits on exposure, capping the number of open positions per instrument and invoking safety stops if drawdowns exceed predefined thresholds, thereby instilling discipline in live trading.
Monitoring and Analysis: For live operation, AlphaFX provides a real-time console dashboard displaying active positions, recent trades, and performance stats. Every trade is logged in detail (to CSV/JSON), enabling thorough post-trade analysis. The same architecture can run in simulation mode for research: it includes an offline backtesting module that replays historical data through the strategy, plus a performance analyzer that reports key metrics (Sharpe ratio, profit factor, win rate, drawdown curves, etc.) to evaluate strategy behavior under varied conditions.

AlphaFX was developed from the ground up as a demonstration of Amin’s expertise in research-informed signal engineering and real-time trading architecture. Early simulations of this system have shown impressive results, for example, in one 5-day test run, AlphaFX achieved a +40% return with an 80% win rate, and it successfully adapted to both low-frequency and high-frequency trading styles by dynamically tuning its signals and risk exposure. These outcomes highlight the platform’s potential and reflect Amin’s rigorous approach to validating ideas under realistic market constraints. Through AlphaFX, he showcases how a holistic integration of data science, software engineering, and domain knowledge can yield a robust trading system that is both innovative and reliable.

Backtest Results

AlphaFX has been tested across multiple scenarios using the strategy_tester.py module and live historical OHLC data. These tests validate end-to-end integration of market ingestion, technical indicator generation, signal alignment (CEP), and live-execution logic under realistic constraints.

Test 1: Low-Frequency Setup (5 Trades, 5 Days)

Instruments: EUR/USD, AUD/USD, USD/CAD
Timeframes: M1, M5, M15
Risk per Trade: 0.8%
Total Trades: 5
Duration: 5 days
Execution Style: High-conviction, low-frequency
Initial Balance: $15,000
Final Balance: $21,066
Net Profit: $6,066
Return: +40.4%
Win Rate: 80%
Sharpe Ratio: 2.21
Signal Efficiency: 34.4%

Test 2: Higher Trade Frequency (11 Trades, 3 Days)

Instruments: EUR/USD, USD/CAD
Timeframes: M1, M5, M15
Risk per Trade: 2.5%
Total Trades: 11
Duration: 3 days
Execution Style: Higher-frequency, aggressive sizing
Initial Balance: $10,000
Final Balance: $14,925
Net Profit: $4,925
Return: +49.2%
Win Rate: 100%
Sharpe Ratio: 4.58
Signal Efficiency: 96.1%

Observations

The system successfully adapted to both conservative and aggressive configurations, maintaining consistent performance with different trade frequencies and risk levels.
CEP logic and multi-timeframe alignment demonstrated robustness across both setups.
Further testing over longer windows and varied market conditions is planned for final validation.

In sum, Sharifi’s contributions to machine learning and quantitative research are marked by an inventive yet practical spirit. Whether designing an autonomous trading agent that learns from scratch (QuantRL) or engineering a full-stack trading system for live markets (AlphaFX), he approaches problems with a blend of scientific curiosity and engineering discipline. This dual expertise in cutting-edge AI algorithms and real-world financial systems enables him to push boundaries at the intersection of art and technology, echoing the same fascination with patterns and complexity that also fuels his musical compositions.

Methodology and Validation

Across projects, Sharifi treats nonstationarity as a first-class constraint. Representation learning supports regime discovery, attention and state-space modeling capture long-horizon dependencies, and uncertainty is addressed with ensembling and stochastic inference. Validation is leak-free by design through walk-forward analysis and purged and embargoed folds. Meta-labeling with the triple-barrier method separates entry logic from outcome conditioning, and transaction cost and capacity models are integrated into objectives and reporting. Documentation, typed interfaces, tests around execution-critical code, and reproducible configurations ensure that research remains auditable and production-oriented.

Notes and Disclaimer

Educational materials only. Nothing on this page constitutes investment advice. Past performance is not indicative of future results.