Axiom Tilt

Research question

We tested whether SEC filing text improves stock selection, then whether a learned allocator improves returns once the selection signal is fixed.

Method

We continued pretraining FinBERT on approximately 227,000 SEC filings, reduced the document embeddings with PCA, and combined them with market, fundamental, and macroeconomic data. A LightGBM LambdaRank model selects thirty stocks each week. A PPO agent assigns portfolio weights under long-only, fully invested constraints, with a linear transaction-cost model.

Findings

Text features improved the after-cost Sharpe ratios of all three tested allocators against their no-text versions, although only the PPO comparison was individually statistically significant. PPO showed no reliable improvement over the deterministic score-proportional rule across three random seeds, and all three seeds underperformed that rule on the untouched 2025 holdout. The equity-only strategies showed no statistically significant risk-adjusted improvement over SPY. A 60/40 blend of the selected stocks and a Treasury ETF had the highest Sharpe ratio in the 2009–2024 comparison.

Limitations

The main walk-forward evaluation covers 2009–2024, with an additional 2025 holdout. The PPO configuration and original seed were selected using 2009, so the paper also reports results excluding that year. Fundamentals coverage excludes roughly 300 historical constituents, execution costs are simplified, and the base-versus-continued-pretrained FinBERT comparison covers only one walk. These historical backtests do not establish expected live returns.

Backtest results

Table 8 compares the deterministic rule with three PPO training seeds. Sharpe ratios include a 5 basis point cost per dollar traded and use the paper's zero risk-free-rate convention.

The 2010–2024 window excludes the year used to select the PPO configuration and original seed. The 2025 holdout was not used for selection. Read the full methodology and limitations ↗