LLM Small-Cap Trading: Pure Beta Outperforms Alpha-Beta Intersection, Sharpe 2.33
This paper presents an end-to-end LLM-driven trading framework for Russell 2000 stocks that decomposes prediction uncertainty into aleatoric and epistemic components, feeds them directly into portfolio covariance matrices, and shows that separating pure alpha and pure beta stock selection triggers outperforms their intersection, with the best configuration (pure beta, GPT-4o mini, Student-t, 40-day hold, risk parity) achieving a Sharpe ratio of 2.33 at 100 bps transaction costs.
Abstract
Large language models extract richer signals from financial news than fixed sentiment dictionaries. This paper investigates an uncertainty-aware construction method : the model's predicted risk — decomposed into aleatoric and epistemic uncertainty — is fed directly into the portfolio allocator's covariance matrix, rather than treating portfolio risk as a fixed historical estimate or only adjusting expected returns.
The full pipeline is evaluated on Russell 2000 stocks, comparing three stock-selection mechanisms: pure alpha trigger (isolates firm-specific anomalies unexplained by macro indicators), pure beta trigger (captures macro-indicator moves before the stock itself reacts), and beta trigger (both channels fire simultaneously). Across the entire holding-period grid, the separated pure alpha and pure beta legs typically outperform the beta intersection in Sharpe ratio and return.
The strongest conservative configuration — pure beta trigger + GPT-4o mini sentiment model + Student-t target distribution + 40-day holding period + risk-parity allocation — achieves a Sharpe ratio of 2.33 at 100 bps transaction costs. The core conclusion: the choice of stock-selection mechanism and allocator is at least as important as the sentiment model , and separating firm-specific and macro risk-exposure triggers carries more information value than requiring both to fire together.
Background
Financial news drives market moves, but fixed sentiment dictionaries miss negation, context, and domain-specific language. LLMs solve this by conditioning on full news context, and recent evidence shows LLM-derived financial sentiment contains return-predictive information superior to dictionary-based measures.
The authors identify two gaps in existing literature. First, most studies score isolated headlines rather than linked news stories — repeated coverage of the same event across outlets disproportionately amplifies a single event's impact . Second, sentiment is typically used only as an expected-return input while portfolio risk remains a historical, signal-agnostic estimate. This is limiting because a predictive model can also provide information about uncertainty and cross-asset covariance.
Related work: Lopez-Lira and Tang (2026) show GPT-4 headline scoring predicts initial reaction and subsequent drift, especially for small caps and negative news. Kirtac and Germano (2024) compare OPT, BERT, FinBERT, and Loughran–McDonald dictionary; OPT long-short achieves Sharpe 3.05 at 10 bps cost. Sentiment-driven portfolio papers usually inject language information into expected returns or Black–Litterman views: Colasanto et al. (2022) use BERT sentiment scores as Black–Litterman views; Lee et al. (2025) transform LLM predictions and prediction uncertainty into views and confidence; Spears et al. (2020) use prediction uncertainty for position sizing but only at the single-holding level, not cross-asset portfolio covariance.
Validation challenges: LLM trading studies face look-ahead bias because training data may cover the evaluation period. Glasserman and Lin (2024) distinguish look-ahead bias from "contamination effects"; Eliseev and Seleznev (2026) use "fake dates" and find modern LLMs struggle to pass the strictest look-ahead screens in macro forecasting; Rizvani et al. (2026) show LLM trading pipelines can be vulnerable to adversarially manipulated headlines. These findings motivate the authors' use of knowledge-cutoff-constrained sentiment scoring , strict forward-looking standardization, and representative story construction.
Problem Statement
The remaining gap is not "whether text contains return information" but how to jointly test the combined contribution of news-summary sentiment, macro/technical stock selection, prediction uncertainty, and asset allocation in a controlled benchmark . The central question: what ultimately determines the performance of a complete LLM sentiment trading pipeline — the sentiment model backend? The return target distribution? The allocator? The holding period? Transaction costs? Or the stock-selection mechanism?
To answer this, the authors build an end-to-end pipeline: from Russell 2000 price and news panels, one of three macro beta triggers defines the tradable cross-section; relevant articles are clustered into news-summary sentiment features; a return-distribution network is trained to output both expected returns and a predictive covariance matrix ; the same predictive moments are fed to six allocation rules, isolating each component's contribution under controlled conditions.
Method
3.1 Stock Selection: Decomposing Macro Beta Triggers
All three mechanisms start from the same Russell 2000 pool, defining a stock-side event set S<sub>S</sub> and an indicator-side event set S<sub>I</sub>. A stock's own anomaly is flagged by a return z-score: Z = (R − μ) / σ The stock's relationship to an indicator is measured by beta: β = Cov(R_i, R_ind) / Var(R_ind) Computed on a rolling window over 58 exogenous indicators (50 Yahoo Finance market series + 8 FRED macro releases). Stock-side trigger requires |Z_i| ≥ 2; indicator-side trigger requires |Z_ind| ≥ 2, absolute beta > 1, and conditional tail beta > 1.
The three selection legs are defined as set operations:
Pure alpha (S<sub>S</sub> \ S<sub>I</sub>) — stock's own anomaly with no concurrent macro trigger to explain it.
Pure beta (S<sub>I</sub> \ S<sub>S</sub>) — macro/market indicator has moved and the stock is exposed, but the stock itself has not yet anomalously moved (anticipatory).
Beta (S<sub>S</sub> ∩ S<sub>I</sub>) — both channels fire simultaneously with consistent direction (confirmatory).
Under each mechanism, candidate stocks are ranked by in-sample BUY events and capped at 50 for model training and backtesting. At allocation time, the filter acts as a mask: the predictive model still outputs full return and covariance predictions, but the allocator may only hold stocks with an active BUY signal under the chosen branch.
3.2 From News to Sentiment: Cluster Stories First, Then Score
Unlike per-article scoring literature, this paper treats sentiment as a property of the "story" not the "article" . For each article–asset pair, the sentiment model returns a calibrated probability distribution over negative/neutral/positive (p<sub>− , p<sub>0</sub>, p<sub>+ ), with raw directional score s<sub>raw</sub> = p<sub>+ − p<sub>− . This undergoes entity-prior correction (subtracting ticker naming/masking bias δ_i), strict trailing shrinkage location estimation (ticker and group median weighted by shrinkage constant κ = 20), and group-day cross-sectional standardization to yield the final standardized sentiment surprise s<sub>final</sub>.
Duplicate coverage is collapsed before entering the model: within a trailing 30-day window, article summaries are embedded and agglomeratively clustered with single linkage; cosine similarity ≥ 0.90 merges into the same story . Each story cluster retains only the article closest to the centroid as the representative news item; the rest are treated as duplicate coverage. Daily sentiment features are the average of these representative articles' final scores, accompanied by a log-scaled story count log(1+K) as a coverage-intensity feature. Neither feature is forward-filled; signals never reflect information unavailable at the moment of use.
Sentiment scoring uses GPT-4o mini, whose October 2023 knowledge cutoff precedes the evaluation period, directly limiting look-ahead bias. To verify results are not artifacts of a specific scorer, three alternative scorers are also tested: FinBERT, Mistral 7B Instruct, and Llama 2 13B Chat.
3.3 Joint Return and Risk Prediction: Multimodal Network
Sentiment signals enter a multimodal network alongside daily price/volume technical features. The news branch uses news-summary sentiment and coverage-intensity features; the daily branch uses OHLCV-derived indicators (close return, overnight return, 20-day rolling volatility, RSI, ATR, Amihud illiquidity, momentum, etc.). Both branches are encoded separately and concatenated over a lookback window to form a per-stock representation h<sub>i,t</sub>.
Unlike prior work that treats risk as a separate historical estimate, this paper predicts both return and risk from the same representation and distinguishes two uncertainty sources. A 0.2 dropout is applied to h<sub>i,t</sub> to obtain stochastic representations; the mean head and covariance head run on each stochastic sample. The covariance head is parameterized as a rank-2 low-rank plus diagonal matrix to structurally guarantee a valid estimate: Σ_A_t,n = F_t,n F_t,n^T + diag(d_t,n) This aleatoric term captures the market's own conditional randomness. Epistemic uncertainty is measured by keeping dropout active at inference and computing the dispersion over S stochastic forward passes:
μ̂_t = (1/S) · Σ_n μ̂_t,n, Σ̂_A_t = (1/S) · Σ_n Σ̂_A_t,nΣ̂_E_t = (1/(S−1)) · Σ_n (μ̂_t,n − μ̂_t)(μ̂_t,n − μ̂_t)^TThe two are combined into a single predictive covariance: Σ̂_t = Σ̂_A_t + Σ̂_E_t This extends the decomposition from a single scalar to a full N×N cross-asset covariance matrix. The model is trained by maximizing the likelihood of realized returns under its own predicted aleatoric distribution (Gaussian or Student-t with fixed degrees of freedom ν = 5), jointly learning risk and return . The network is a 1D convolutional encoder (32 channels, kernel size 3) + shared 64-dimensional hidden representation, AdamW optimizer, batch size 32, learning rate 1e-3, weight decay 1e-5, gradient clipping 1.0, early stopping on validation negative log-likelihood. Before the portfolio stage, μ̂_t and Σ̂_t are converted from log-return moments to arithmetic-return moments.
3.4 Risk-Aware Portfolio Construction
The combined covariance is fed directly into portfolio construction. Unlike prior sentiment-driven work that only adjusts expected-return inputs while keeping the covariance matrix historical, this paper feeds Σ_t into a standard mean-variance optimizer with transaction-cost regularization, full investment, 40% single-asset cap, and risk-aversion coefficient δ = 2.5. All reported filtered portfolios are long-only. The optimizer is penalized more heavily when the model itself is uncertain, not merely when assets are historically volatile.
As controls, five standard alternatives are benchmarked: equal weight, risk parity (RP), hierarchical risk parity (HRP), Black–Litterman, and Bayesian Black–Litterman. Each baseline consumes the same (μ_t, Σ_t) for a controlled comparison, isolating the allocation rule's effect from the predictive moments themselves.
Experiments
Experimental Setup
Dataset : Russell 2000 stocks. Daily OHLCV from 2023-10-02 to 2025-12-31; scored news from 2023-10-01 to 2025-12-31; macro beta selection additionally uses macro indicator panel from 2022-01-01 to 2026-01-01. All selection, standardization, model fitting, and early stopping use only the in-sample period up to 2024-12-31 (first 80% trading days for training, last 20% for validation, split boundary has holding-period embargo). Calendar year 2025 is held out for out-of-sample backtest.
Benchmark grid : 3 selection mechanisms × 4 sentiment backends × 2 target distributions (Gaussian/Student-t) × 8 holding periods (1, 2, 3, 5, 10, 20, 40, 60 trading days) × 8 cost levels (0–100 bps plus fixed 5 bps one-way slippage) × 6 allocation methods. Performance reported via cumulative net return, annualized return, annualized volatility, Sharpe, Sortino, Calmar, max drawdown, average turnover, total cost, hit rate, and exposure.
Stock Selection Mechanism & Holding Period
At the 100 bps conservative benchmark, separated pure alpha and pure beta legs are more useful than the beta intersection : pure alpha leads at 5, 10, and 60 trading days; pure beta leads at 20 and 40 days. The one-day holding period is cost-sensitive: pure beta is strongest at low cost but its advantage vanishes under the 100 bps stress scenario.
This pattern has a direct economic interpretation. Pure alpha selects anomalous stock moves without a concurrent macro trigger — a small-cap slow-information-diffusion scenario: sparse coverage, limited liquidity, constrained arbitrage capacity, so information takes several trading days to fully price. Pure beta selects stocks exposed to an anomalous macro/market indicator move before the stock's own anomaly appears: it can be a one-day lead-lag signal at low cost, or a slower 40-day macro transmission signal — investors gradually map rate, sector, commodity, volatility, and demand shocks into heterogeneous small-cap fundamentals.
Optimal Configuration & Cost Robustness
Strongest conservative row : pure beta + GPT-4o mini sentiment + Student-t target + 40-day holding + risk-parity allocation, achieving Sharpe 2.33 at 100 bps cost (cumulative net return 100.1%, annualized return 95.9%, max drawdown −18.3%); Sharpe 2.51 at 20 bps, 2.44 at 50 bps. At 60-day holding, pure alpha is strongest across the entire 0–100 bps range (FinBERT + Gaussian + HRP, Sharpe 1.96–2.10), and its lead actually widens as cost rises — because 60-day rebalancing is inherently low-turnover, the cost scan only mildly shifts absolute Sharpe levels but rules out the possibility that pure alpha's edge is an artifact of a particular cost assumption.
Holding Period Comparison & Mechanism Interpretation
Holding-period comparison is non-monotonic because the two separated legs capture different adjustment speeds. The one-day pure beta is a clear lead-lag trade: the signal starts from already-occurred liquid macro/sector/rate/commodity/volatility/index moves, buying small caps exposed to them but not yet anomalously moved; but it is fragile to daily rebalancing repeatedly paying spreads and opening noise, so its advantage disappears at 100 bps cost. The 20- and 40-day pure beta stems from slower macro repricing, as investors translate aggregate shocks into heterogeneous small-cap cash-flow and discount-rate effects.
Pure alpha is more credible at 5, 10, and 60 days; these trades start from firm-specific anomalies without a concurrent macro trigger, so the information is more likely company-level: earnings revisions, financing news, contracts, governance changes, litigation, product updates, etc. Small-cap analyst coverage is thin and arbitrage capacity weak, so information can diffuse slowly over several trading days. The 60-day result aligns with a small set of persistent firm events — low turnover lets fundamental repricing dominate trading friction.
The beta intersection is weaker overall because it simultaneously requires both channels to fire, precisely dropping the two useful cases: early macro spillover before the stock reacts, and idiosyncratic firm events without a broad factor shock. The separated legs therefore resemble distinct trading hypotheses rather than minor variants of the same filter: pure beta is a lead-lag or macro-transmission strategy; pure alpha is a delayed firm-specific repricing strategy — the preferred leg shifts with holding period and transaction cost, rather than a universally dominant selection rule existing.
Signed-in readers can open the original source through BestHub's protected redirect.
This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactand we will review it promptly.
How this landed with the community
Was this worth your time?
0 Comments
Thoughtful readers leave field notes, pushback, and hard-won operational detail here.
