Technical Summary: Semantic State Abstraction Interfaces for LLM-Augmented Portfolio Decisions

Problem Statement

Deep Reinforcement Learning (DRL) and sequential decision systems increasingly integrate unstructured text (e.g., financial news) with numerical market data. However, a critical design gap exists: most pipelines obscure the distinction between the semantic representation of text and the optimization algorithm used to exploit it. Common approaches either use high-dimensional latent vectors (obscuring interpretability) or single scalar scores (e.g., a single sentiment score), which conflate orthogonal concepts like risk, confidence, and volatility. Furthermore, when text is sparse (appearing only on a subset of decision dates), it is unclear whether performance gains stem from the semantic signal itself or from the specific learning algorithm's ability to handle noise.

Methodology: Semantic State Abstraction Interfaces (SSAI)

The authors introduce Semantic State Abstraction Interfaces (SSAI), a methodological template designed to map sparse, unstructured text into $K$ auditable, named coordinates.

1. The SSAI Framework

• Definition: SSAI maps a set of documents $N_{s,d}$ for asset $s$ on day $d$ to a vector $\sigma \in \{1, \dots, 5\}^K$.
• Axes: In this instantiation, $K=4$ axes are defined:
  1. Sentiment: Very negative (1) to very positive (5).
  2. Risk: Very low risk (1) to very high risk (5).
  3. Confidence: Very uncertain (1) to very certain (5).
  4. Volatility Forecast: Very calm (1) to highly volatile (5).
• Implementation: A frozen Large Language Model (LLM) is prompted with fixed instructions and few-shot examples to extract these integer scores.
• Sparsity Handling: On days with no news ($N_{s,d} = \emptyset$), the system applies a neutral default (value 3.0) rather than zero-filling or dropping the data.
• Auditability: The interface is designed for human-in-the-loop oversight, allowing for prompt-level ablation and regulatory legibility (e.g., "risk increased from 2 to 4").

2. Evaluation Protocol: Separating Representation from Optimization

The core methodological contribution is an evaluation design that isolates the effect of the representation ($\phi$) from the optimizer.

• Fixed Representation: A single, frozen $\phi$ (the SSAI map) is shared across all estimators.
• Parallel Estimators: The same semantic state is fed into three distinct classes of models:
  1. Low-variance estimators: Direct factor portfolios (SFP, SRF, SCW) and supervised ridge forecasters.
  2. High-variance nonlinear estimators: Reinforcement Learning agents (DP-PPO and SAC).
• Invariant: Because $\phi$ is identical and frozen across all models, any performance differences are attributable to the algorithm choice, not the representation.

3. Experimental Setup

• Data: 30 NASDAQ-100 equities, 2019–2023 test period (out-of-sample), trained on 2013–2018. News sourced from FNSPID.
• Baselines: Buy-and-Hold (B&H), Momentum, Equal-Volatility, and dense neural encoders (FinBERT, PC1 of SSAI axes).
• RL Agents:
  • DP-PPO: Drawdown-penalized Proximal Policy Optimization with a reward shaping term to penalize drawdowns super-linearly.
  • SAC: Soft Actor-Critic (off-policy) with identical state observations.

Key Results

1. Portfolio Performance (Factor Portfolios)

• Apparent Outperformance: The 4-axis SSAI Semantic Factor Portfolio (SFP) achieved a Cumulative Return (CR) of 307.2% and Sharpe of 1.067, seemingly outperforming Buy-and-Hold (243.6%).
• The "Composition Effect" Caveat: This advantage does not survive rigorous controls:
  • Stratum-Matched Failure: When controlling for news coverage frequency, SFP underperformed B&H in every coverage tercile (Low, Mid, High). The aggregate gain was a "basket-selection composition effect"—the model selected mid-coverage stocks that happened to perform well, not because the signal was predictive within those groups.
  • Cost Sensitivity: The advantage reverses at transaction costs $\ge 0.2\%$.
  • Statistical Significance: The difference between the 4-factor SFP and a sentiment-only SFP was non-significant (Wilcoxon $p=0.556$).
• Superior Alternatives:
  • PC1-SFP: A data-driven first principal component of the four axes achieved 433.6% CR (126 percentage points higher than SSAI-SFP).
  • FinBERT-SFP: A dense encoder baseline achieved 386.3% CR.
  • Conclusion: The named-axis interface (SSAI) incurs a significant performance cost (126pp CR) relative to data-driven or dense alternatives in this setting.

2. Supervised Forecasting

• Naively appending sparse LLM semantic columns to price/technical features hurt ridge forecasting performance.
• Lexical baselines (VADER, TF-IDF/SVD) and "high-conviction semantic tilts" recovered performance to near price-only levels, but dense neural encoders (FinBERT) did not improve ridge forecasting without careful integration.

3. Reinforcement Learning Diagnostics

• Algorithm Dependence: Under the identical SSAI state, SAC (Sharpe 1.059) significantly outperformed DP-PPO (Sharpe 0.920, $p=0.027$).
• Interpretation: This gap is attributed to algorithmic differences (off-policy replay, entropy regularization) rather than representation quality.
• Masking: Removing LLM signals (neutral masking) did not significantly degrade DP-PPO performance (Wilcoxon $p \approx 0.25–0.31$), suggesting the RL agents were not effectively utilizing the sparse semantic signal.
• Ablation: Removing specific axes (e.g., confidence) reduced returns, but the single-checkpoint reversal (neutral > full signals) was attributed to evaluation-time state distribution shifts, not signal harm.

Significance and Claims

The paper explicitly frames its contribution as diagnostic and cautionary rather than a claim of superior performance.

1. SSAI as an Interpretability-Performance Frontier Instrument: The study quantifies the "cost of interpretability." By comparing SSAI (auditable, named axes) against PC1 (data-driven) and FinBERT (dense), the authors show that maintaining auditable, human-readable axes comes at a substantial performance cost (126pp CR over five years) in this specific portfolio context.
2. Separation of Signal and Optimizer: The SSAI template provides a reusable method to disentangle whether a system's success is due to the semantic representation or the learning algorithm. The results indicate that in this setting, algorithm choice (SAC vs. PPO) dominates representation effects.
3. Cautionary Findings on Sparse Text:
   • Sparse LLM signals are useful as high-conviction tilts or for risk management but are not robust as naive inputs for dense forecasting or RL without careful integration.
   • Apparent outperformance in factor portfolios can be a composition artifact (selecting specific stocks) rather than a true signal advantage.
   • High-coverage stocks (e.g., NVDA, GOOGL) showed the worst relative performance for SSAI, likely due to market efficiency and factor model misspecification during the AI investment cycle.

Conclusion: The authors do not claim SSAI outperforms denser alternatives. Instead, they offer SSAI as a methodological template for future research to rigorously separate semantic signal from optimization noise, while highlighting that the current trade-off between auditability and predictive power is significant. The work serves as a "retrospective benchmark" emphasizing that interpretability constraints have measurable economic costs.