QUANTA: A Regime-Filtered Hybrid Trading Agent Comparing LSTM and LightGBM Architectures

The exponential digitization of global financial markets has necessitated a fundamental paradigm shift in retail trading strategies, transitioning from manual, discretionary execution to automated, data-driven systems. This paper presents the comprehensive design, implementation, and rigorous evaluation of QUANTA (Quantitative Unified Autonomous Neural Trading Agent), a novel cloud-based algorithmic trading infrastructure. We address the critical challenge of non-stationarity and high noise-to-signal ratios in cryptocurrency markets by investigating the efficacy of two distinct machine learning architectures: Long Short-Term Memory (LSTM) networks, representing deep learning, and Light Gradient Boosting Machine (LightGBM), representing ensemble tree-based learning. Utilizing a high-fidelity historical dataset of Bitcoin (BTC/USDT) sourced from the Binance exchange spanning an eight-year period (2017-2025), we implement a novel "Smart Money Concept" (SMC) feature engineering pipeline coupled with a strict logicbased regime filter. Our empirical backtesting results reveal a significant divergence in model performance under identical market conditions. While LSTM models are theoretically superior for sequence modeling, they exhibited a marked tendency to overfit stochastic noise in tabular technical data. Conversely, the LightGBM architecture demonstrated superior generalization, stability, and execution speed. The proposed system achieved a Net PnL of 310.55% ($77.64) compared to the LSTM baseline of 201.25% ($50.31) on a starting micro-capital of $25. Furthermore, to validate the operational feasibility of the selected model, the system was deployed on a live execution environment via the Delta Exchange Testnet, orchestrated by a robust cloud-native architecture using AWS EC2 and Google Firebase.