0% Complete
صفحه اصلی
/
دومین همایش بین المللی هوش مصنوعی
Robust Algorithmic Trading in Volatile Markets Using Pessimistic Transformer-Enhanced TD3 in Continuous Action Spaces
نویسندگان :
Amirhossein Ghozati
1
Armin Salimi-Badr
2
1- دانشگاه شهید بهشتی
2- دانشگاه شهید بهشتی
کلمات کلیدی :
Deep Reinforcement Learning،Algorithmic Trading،TD3،Pessimism،Transformer Attention،Multi-Asset Portfolio،Continuous Actions
چکیده :
Contemporary financial markets, marked by their rapid evolution and inherent volatility, pose substantial challenges to algorithmic trading due to pervasive uncertainty. Traditional approaches frequently prove insufficient, and even advanced deep reinforcement learning (DRL) models may succumb to overestimation biases. In response, we present the Pessimistic Transformer-based TD3 (PT-TD3), an innovative extension of the Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithm, specifically adapted for continuous action spaces. Our methodology integrates a Transformer encoder to discern long-term patterns and interdependencies among multiple assets, alongside a pessimism mechanism that exploits variances between twin critics to conservatively modulate Q-values, thereby curtailing unwarranted overconfidence in risky decisions. Additionally, we enhance the state representation with technical indicators such as RSI and MACD, while refining the reward function through volatility penalties to promote greater stability. Evaluated on a diversified portfolio of 10 U.S. stocks over the period from 2015 to 2023, PT-TD3 yields a total return of 52.14% (exceeding the benchmark's 46.03%), a Sharpe ratio of 3.63 (versus 2.76), and a maximum drawdown of -5.06% (compared to -7.95%), surpassing competing techniques including LSTM-PPO hybrids and CNN-GRU-Attention frameworks. An ablation study reveals optimal performance at , underscoring how this calibration effectively manages uncertainty without compromising potential gains. Ultimately, PT-TD3 furnishes a more resilient framework for traders operating in unpredictable market conditions.
لیست مقالات
لیست مقالات بایگانی شده
Strategies and Future Horizons of Innovative Entrepreneurship in AI-Based Programming
Milad Ghiasspour
Attention-Based Noise Reduction for Surface-Electromyography: A Novel Method for Enhanced Signal Quality in Clinical Diagnostics
Seyyed Ali Zendehbad - Abdollah PourMottaghi - Marzieh Allami Sanjani
Hybrid Deep Learning Models for Cardiovascular Disease Prediction: A Comprehensive Review of Convolution-Transformer Architectures
Ali Azimi Lamir - Masoud Bekravi - Babak Nouri Moghaddam
Enhancing Drug Synergy Prediction through Relation and Structure Aware Attention Networks
Dorsa Abdi - Abdussalam Nik Kerdar - Parinaz Kanan - Amirali Maroufi - Maedeh Sadat Tahaei
Title Generation for the Qur'anic chapters by summarizing them
Masoume Maleki - Alireza Talebpour - Mostafa Moradi
Deep Learning Frailty Model for Heart Failure Survival Prediction
Solmaz Norouzi - Mohammad Asghari Jafarabadi - Ebrahim Hajizadeh - Hossein Khormaei - Nasim Naderi
Automated Recognition of Marine Thermal Patterns Using Deep Learning
Alireza Sharifi - Alireza Vafaeinejad
Customer Segmentation in Online Tire Sales Using RFM and Quantity: A Comparison of K-Means, MiniBatchKMeans, and DBSCAN with Quality Improvement via Density-Based Refinement
Seyed Mohammadreza Jalalian Shahri - Seyed Hashem Mohtashami
Intermediate Fine-Tuning for Robust Persian Emotion Detection in Text
Morteza Mahdavi Mortazavi - Mehrnoush Shamsfard
Efficient and Accurate Fairness Verification for Quantum Variational Circuits
Sajjad Hashemian Meymandi - Mohammad Saeed Arvenaghi
بیشتر
ثمین همایش، سامانه مدیریت کنفرانس ها و جشنواره ها - نگارش 44.5.0