Financial Numerical Prediction and Allocation as Token Generation

Imagine you're trying to predict stock returns or allocate investments in a way that feels intuitive and direct. Traditionally, financial prediction relies on separate models for regression or ranking, which can lead to disjointed decision-making and missed opportunities. This separation can create issues like misalignment between the model's outputs and the actual financial decisions that need to be made, which is what's called model misalignment. It often results in suboptimal performance when the market dynamics change or when faced with new data.

To address this, the authors propose FinATOM, a unified approach that allows a causal language model to generate forecasts and decisions directly through token generation. Instead of relying on separate heads for different tasks, this model emits volatility-standardized return tokens and uses a novel training process that includes ordinal and ranking supervision. The allocation model then generates normalized weights for investments, optimizing for performance while maintaining consistency with a mean-variance anchor. The results show that this method improves Sharpe ratios significantly in real-world ETF tests, suggesting that direct token generation can enhance financial decision-making and forecasting. For builders, this means a more integrated and potentially more effective way to leverage language models in finance, moving beyond traditional methods.

The approach of using a causal language model for direct financial predictions is a meaningful extension of existing methods.

The results are supported by rigorous testing across multiple metrics and scenarios.