Systematic strategies based on factors like value, momentum, quality, etc.
An explainable multi-agent system using fine-tuned GPT-4o models for crypto portfolio management. Specialized agents analyze news, factors, and charts, collaborate on decisions, and execute trades—outperforming benchmarks in returns, accuracy, and interpretability.
This paper shows that large language models (LLMs) can predict stock returns in China using public news. The best models—especially an ensemble of multiple LLMs and Baichuan—produce daily trading signals that generate up to 91% annualized alpha, proving highly effective in emerging markets.
GPT-4 can create alpha. By simply prompting it with price and volume definitions, it generates factors that deliver Sharpe ratios up to 4.49 and annual returns up to 66%—without using any financial data.
This paper introduces asset embeddings—vector representations of stocks learned from investor portfolios using techniques like BERT and Word2Vec. These embeddings outperform traditional characteristics in explaining return comovement, offering a new framework for understanding investor behavior.
ChatGPT can extract forward-looking investment signals from earnings calls. The paper shows that its investment score predicts firm capex up to 9 quarters ahead and negatively forecasts future stock returns—revealing information not yet priced by markets.
This paper shows that a simple, model-free valuation approach using all available accounting data can identify stock mispricing. A strategy that buys underpriced and sells overpriced firms earns **4–10% alpha per year**, outperforming many known anomalies—even without complex valuation models.
Short sellers face unique risks, including loan recalls and rising borrow fees. This paper shows that higher short-selling risk predicts **lower future returns**, **more mispricing**, and **less efficient prices**. These effects are stronger for trades with long expected horizons.
This paper compares machine learning methods to predict stock returns and measure equity risk premia. The best-performing model (a 3-layer neural net) earns out-of-sample Sharpe ratios as high as 2.45.
This paper introduces a new class of asset pricing models using autoencoders. By embedding firm characteristics into a neural network architecture, the authors construct a nonlinear conditional factor model that enforces no-arbitrage. It dramatically outperforms linear benchmarks.
Machine learning (ML) signals look great in backtests—but they often break when exposed to economic restrictions. This paper shows that ML-based strategies lose 50–80% of their alpha when excluding microcaps, distressed stocks, or accounting for transaction costs.
This paper uses deep learning to estimate the stochastic discount factor (SDF) that prices all U.S. stocks. The authors combine a feedforward network, LSTM, and a generative adversarial network (GAN) to build a non-linear, no-arbitrage-compliant model.
Multinational firms' returns are predictably linked to foreign industry news from their sales regions. This paper shows that investors underreact to economically relevant information abroad.