🤖 AI量化金融论文

We propose a novel deep reinforcement learning framework for portfolio optimization that dynamically adjusts risk constraints based on market volatility. Our approach utilizes a hierarchical policy gradient method that separately learns asset allocation and risk management policies. Experimental results on S&P 500 constituents demonstrate superior risk-adjusted returns compared to traditional mean-variance optimization and existing RL methods.

This paper introduces a Temporal Fusion Transformer (TFT) based approach for multi-horizon stock return forecasting. We demonstrate that TFT's interpretable attention mechanisms can identify regime changes in financial markets. Our model outperforms existing deep learning methods on both short-term (1-5 days) and long-term (1-12 months) horizons.

We present an LSTM-based system that generates real-time trading signals by combining financial news sentiment with technical indicators. Our model processes both textual news data and time-series price data through separate LSTM encoders, then fuses them for prediction. Backtesting shows an annual return of 18.7% with a Sharpe ratio of 1.34.

We propose a hierarchical reinforcement learning framework that adapts trading strategies to varying market regimes. The high-level controller determines market regime and sets strategic objectives, while the low-level controller executes specific trading actions. Our approach significantly reduces maximum drawdown while maintaining competitive returns.

We introduce a novel graph neural network architecture for modeling complex dependencies across multiple financial assets. Our approach constructs a dynamic graph where nodes represent assets and edges capture correlations. Experiments on 50 liquid assets show 23% improvement in risk-adjusted returns.

We adapt the Informer model for financial time-series forecasting, addressing the challenge of long-term prediction in stock markets. Our modified architecture includes a ProbSparse self-attention mechanism that reduces computational complexity from O(L²) to O(L log L). We demonstrate improved forecast accuracy for 30-day and 90-day stock price predictions.

We develop an attention-based deep learning model to predict order book imbalance and its impact on short-term price movements. By incorporating spatial attention across price levels and temporal attention across time, our model captures both the vertical structure and dynamic evolution of the order book.

We present FinGPT, a large language model specifically tuned for financial market analysis. FinGPT outperforms existing models on financial sentiment analysis, news summarization, and earnings prediction tasks. We also introduce a new dataset for financial NLP research.