Application of Artificial Intelligence Algorithms in Gain Optimization Problems

Abstract

The cryptocurrency market is volatile, presenting both a unique challenge and an opportunity for algorithmic traders. This study evaluates six algorithmic trading models — Deep Q-Network (DQN), Proximal Policy Optimization (PPO), Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), Random Forest (RF), and XGBoost — across five assets (Bitcoin, Ethereum, Cardano, Solana, and Binance Coin) and five investment horizons (10 days, 1 month, 3 months, 6 months, and 12 months), using 2024 out-of-sample market data. The proposed framework addresses two key limitations in prior work: (1) reinforcement learning agents are restricted to a fixed three‑action space without position sizing, and (2) the absence of multi‑horizon evaluation frameworks. The models are evaluated on return, Sharpe ratio, and maximum drawdown. Key quantitative findings include: DQN achieves +36.59% on SOL at 1 month and +49.52% on BNB at 3 months, outperforming all other models by a large margin in trending markets. PPO delivers the highest single‑asset return of +74.60% on ADA at 12 months, demonstrating its ability to learn when to stay out of bearish markets. RF records the tightest drawdown control, with a maximum drawdown of only –2.05% on BNB at 1 month, validating its capital preservation capability. No single model dominates across all assets and horizons; the optimal choice depends on the interplay between market regime, asset behaviour, and investment horizon. All six models are deployed in Trademind IA, a real‑time AI trading dashboard available on mobile and desktop. This work opens a clear path toward a dynamic, regime‑aware model selection framework that automatically assigns the best model to the current market condition. Future work should extend the evaluation to multiple market cycles and incorporate asset‑specific on‑chain features.