Technical Summary: FinATOM

Problem Statement

Financial prediction and decision-making systems typically rely on task-specific regression, ranking, or policy heads attached to language models. This approach fragments the output contract, requiring separate parameterizations, losses, and decoders for each task, thereby separating the language model from the final numerical object being evaluated. Furthermore, standard numerical token generation faces challenges such as irregular tokenization, the loss of numerical order, and the risk of generating malformed outputs that violate constraints (e.g., long-only weights, budget limits). The paper investigates whether a causal language model can directly represent forecasts and decisions through constrained token generation, eliminating the need for task-specific numerical heads while maintaining auditability, mechanical parseability, and compatibility with both supervised fine-tuning (SFT) and token-level policy optimization.

Methodology

FinATOM introduces a unified, head-free interface using a LoRA-adapted Llama 3.2 1B backbone. The system supports two complementary tasks—stock-return forecasting and dynamic ETF allocation—using a shared vocabulary projection but distinct output grammars and training objectives.

1. Shared Token-Native Interface

Instead of appending regression or policy heads, the model generates constrained tokens directly. Valid numerical values are learned vocabulary items scored by the tied language model readout.

• Forecasting: The model autoregressively emits tokens for consecutive D1, D2, and D3 returns.
• Allocation: The model generates normalized, long-only weights for five ETFs.

2. Portfolio Allocation Pipeline

The allocation task is separated into causal imitation and outcome-based improvement:

• Causal SFT (Supervised Fine-Tuning): The model learns to imitate a "causal mean–variance anchor." This teacher target is computed exclusively from information visible at the decision date (20-day returns, risk, correlations, and the previous portfolio state). The continuous solution is shrunk (85% anchor, 15% equal weight) and deterministically quantized onto a 50-unit grid. The SFT objective combines token cross-entropy with an ordinal Gaussian target at the five value slots.
• Policy Optimization (DAPO-Augmented GRPO): A subsequent stage optimizes the realized 21-day Sharpe ratio. The model samples legal allocation sequences, and the reward is defined as the realized Sharpe minus a penalty for deviation from the causal anchor. Future returns enter only this reward signal, not the prompt or SFT targets. The update uses Group Relative Policy Optimization (GRPO) with DAPO-style asymmetric clipping and token-level gradients, without a reference-model KL penalty.

3. Forecasting Pipeline

• Target Construction: Targets are volatility-standardized returns clipped to a range of [-8, 8], mapped to 41 quantile bins.
• Training: SFT uses ordinal cross-entropy on the bucket slots and a same-date ranking loss.
• Decoding: The model generates D1, D2, and D3 tokens autoregressively. Continuous scores are derived from the conditional expectations of the ordered bucket distributions, rescaled by trailing volatility, and compounded for ranking.
• Policy Stage: A one-epoch policy stage samples trajectories and updates the three bucket positions using a multi-component financial reward (step, path, direction, rank, and top-20 components).

Key Contributions

1. Unified Head-Free Interface: Introduction of a shared constrained token interface for autoregressive three-step return forecasting and normalized five-ETF allocation, removing the need for additional task-specific numerical heads.
2. Causal Allocation Pipeline: Development of a pipeline combining a fully specified mean–variance teacher (causal SFT) with DAPO-augmented GRPO, aligning the policy directly with realized portfolio Sharpe while penalizing deviation from the causal anchor.
3. Empirical Evidence: Provision of chronological allocation evidence across three test periods (2023–2025), a controlled input-modality ablation, and an exploratory comparison on the FinTexTS dataset, distinguishing within-protocol improvements from contextual comparisons with published baselines.

Results

Portfolio Allocation (2023–2025)

• Sharpe Improvement: The policy stage improved the pooled gross Sharpe from 1.428 (SFT) to 1.529. Under a 5-basis-point transaction cost model, net Sharpe increased from 1.394 to 1.494.
• Regime Dependence: The policy matched SFT performance in 2023 but showed significant gains in 2024 and 2025. In 2024, the policy Sharpe (1.564) exceeded the causal teacher's Sharpe (1.407), indicating the policy learns beyond simple imitation.
• Turnover: Average one-way turnover increased only slightly from 1.481% to 1.537%, suggesting the gains stem from improved return–risk tradeoffs rather than increased trading intensity.
• Input Modality: The combination of news and time-series data yielded the highest mean Sharpe (1.540). While time-series data provided a stable standalone signal, news data contributed most significantly in the 2025 regime.

Stock Forecasting (FinTexTS)

• Performance: On the FinTexTS dataset, the SFT strategy achieved a cumulative return of 73.52% and a Sharpe of 2.68. The one-epoch policy stage modestly improved these to 73.72% and 2.69, respectively.
• Comparison: While the FinATOM policy outperformed the published FinTexTS system in cumulative return (73.72% vs. 53.98%), the Sharpe difference was minimal (2.69 vs. 2.67). The authors note this comparison is contextual due to differing protocols and lack of matched volatility/drawdown metrics.

Significance and Claims

The paper claims that FinATOM demonstrates the feasibility of using direct language-model token generation for financial numerical prediction and decision-making.

• Feasibility over Superiority: The authors explicitly state that the results establish feasibility rather than universal superiority. The policy optimization provides incremental refinement rather than a wholesale change in the signal, particularly in the forecasting task where most predictive power is captured during SFT.
• Causal Integrity: The design ensures that future returns are absent from prompts and SFT targets, entering only the policy reward, thus maintaining causal integrity.
• Limitations: The study acknowledges limitations, including a restricted asset universe (5 ETFs), a single training seed per setting, and the potential influence of the Llama 3.2 pretraining cutoff (December 2023) on the 2023 results. The authors caution that the findings should be interpreted as evidence of feasibility, warranting broader tests across assets, regimes, and random seeds to confirm robustness.