AlphaRLM's 8B S&P 500 book earns 6.94% annualized excess return with depth-8 XGBoost trees. Keep its selected factors and switch to Ridge regression, the other predictor the paper tests, and it trails the index by 3.90% a year.
That difference governs my reading of Huang, Hong and their coauthors. The US result comes from a selected factor pool, a nonlinear predictor and a daily trading rule. The authors present it as evidence for the factors themselves. Their diagnostics make a stronger case for that interpretation in the Chinese markets than in US large caps.
From abstracts to trades
Qwen3 reads the title and abstract of a finance paper, then writes one executable Qlib expression. The model comes in 8B and 14B parameter versions. It works with six daily fields ($open, $close, $high, $low, $volume, $vwap) and a fixed operator list, with no seed factor to mutate.
A valid expression receives a reward of 1, plus 4 × 100 × TMB. Validity means it parses and its historical values stay fixed after the addition of 1 day, 1 week, 1 month or 1 year of future data. That future-extension test is a sensible, cheap check for look-ahead. TMB measures separation between groups of returns. Each day, the authors sort stocks into 100 factor quantiles, pool next-day returns within each quantile across 2018 and 2019, and subtract the lowest group mean from the highest, regardless of where those groups fall in the ranking. DAPO is the group-relative policy-gradient method used to train the model: it samples 8 completions per prompt and favours those that beat their siblings.
The literature supplies the starting idea. An abstract on price impact, asymmetric volatility or higher moments gives the model an economic prior; market feedback rewards the price-volume expression that works. The ablation supports a role for that prior. When DAPO trains on non-financial text instead of finance abstracts, excess return falls from 12.14% to 2.13% on CSI 300 and from 6.94% to 2.90% on the S&P 500. The corpus contains 4,443 papers published from 2015 to 2019: 1,720 training prompts, 430 validation prompts and 2,293 held-out generation prompts.
After generation, the authors screen the pool for validity and rank it by XGBoost feature importance on 2015 to 2019. They retain 1,000 factors. A second XGBoost, with 600 trees, depth 8 and eta 0.04, predicts rank-normalized next-day returns. The daily book holds the top 50 names while excluding the bottom 5 from its candidate pool. Results cover 2021 to 2024 after 0.05% opening and 0.15% closing costs. Across all three markets, the 14B and 8B variants place first and second for annualized excess return. On the S&P 500, they post 8.08% (±1.25%) and 6.94%, respectively; the strongest baseline, FactorMiner, posts 3.16%. Their drawdowns are -14.77% and -14.84%. The paper annualizes the arithmetic mean of daily excess returns by multiplying it by 252, and measures drawdown on summed rather than compounded excess returns.
What survives a linear predictor?
The S&P 500 correlations barely register. The 8B pool posts an IC of 0.0021, a RankIC of 0.0031 and an ICIR of 0.0158 alongside its 6.94% excess return. The authors acknowledge the mismatch: return advantages do not come with uniformly better IC or RankIC, they write, and marginal predictive correlations alone do not fully account for portfolio performance. They use the Ridge run to argue that "nonlinear interactions learned by XGBoost are not the sole explanation" for the gains.
For the A-share universes, Ridge preserves the AR ranking. The 14B pool remains first, with 8.32% on CSI 300 and 9.58% on CSI 1000. The S&P 500 table tells a harsher story. AlphaRLM-8B leads at -3.90%, ahead of the 14B pool, while every method in the table loses. The authors report the negative figures and claim a relative advantage: this run asks whether their lead depends entirely on XGBoost-specific nonlinear interactions. The lead survives. The positive 6.94% does not survive the predictor change, so the Ridge result offers no positive linear signal from the US factors. First place here means the smallest loss.
On US large caps, the gain therefore appears to depend on interactions among 1,000 weak features in a tree model fitted on 2015 to 2019. The authors also point to a factor-level result: AlphaRLM-8B has the highest median single-factor TMB on the S&P 500 test period, 2021 to 2024. TMB is the training objective, though, and its max-minus-min construction catches sample extremes. A high median offers limited evidence for a tradable linear signal. We did not find style or beta neutralization in the paper, leaving open whether size or short-term reversal drives the result.
The selection inside the result
The language model's output passes through two selections based on noisy in-sample estimates. First, TMB chooses the maximum and minimum among 100 group means. CSI 300 is the A-share universe used to train the policy; with 100 groups, each has roughly 3 stocks per day. The reward pools group observations over 2018 and 2019, so a mean uses two years of stock-days rather than just 3 names. Even with that pooling, selecting the extremes of 100 estimates creates upward bias. The authors derive a finite-sample bound whose estimation error increases with the number of groups. They give the countervailing cost too: coarser grouping "can average away localized non-monotonic structure". For them, the group count trades resolution against error. They call the bound "an interpretive reference rather than a distribution-free guarantee" and TMB "a task-level surrogate" for factor usefulness, short of a full characterization.
The ablations nevertheless favour 100 groups. With 30, excess return falls to 8.53% on CSI 300 and 2.11% on the S&P 500. An endpoint reward, the absolute top-quantile minus bottom-quantile spread, produces 9.68% and 0.77%. Its CSI 300 IC is higher, 0.031 against 0.026, while its drawdown reaches -15.60% against -6.15%.
Next comes the factor count, which the authors say they fixed at 1,000 using training-period information. In the S&P 500 test, 1,500 factors yield 2.27%; 2,000 yield -0.33% and a -32.47% drawdown. CSI 300 peaks at 360 factors, with 12.65% versus 12.14% at 1,000. Extending the same ranked pool takes the US figure from 6.94% below zero.
The training reward settles near 21 in both markets. The policy has learned its objective; how much of that objective reflects noise remains open.
Our run on 300 US large caps
We could not run AlphaRLM itself. We lack the curated 2015 to 2019 literature corpus and the capacity to train 8B or 14B policies with DAPO. We instead searched a deterministic library for classical symbolic expressions, retaining the paper's future-extension validity check. Our reward used an endpoint version of grouped-return separation: 30 groups and a signed top factor-quantile mean minus bottom factor-quantile mean on 2015 to 2019 open-to-open returns. We then froze the sign. This is closer to the paper's endpoint ablation than to TMB. We traded one expression rather than feeding 1,000 to XGBoost, and changed the universe to the 300 largest non-ADR US stocks from an annual capitalization screen instead of historical S&P 500 membership. This adaptation says nothing about AlphaRLM's claimed advantage.
From January 2020 to July 2024, our book compounded at 17.46% a year. Its volatility was 28.33%, its Sharpe ratio 0.83, and its beta to SPY 1.09. Maximum drawdown reached -50.75% across 43,938 trades. The paper gives no Sharpe ratio for comparison. For AlphaRLM-8B over 2021 to 2024, it reports 6.94% excess return and a -14.84% drawdown. Our return is higher, and our drawdown is far deeper. Our figure is an absolute return; theirs is index-relative, from a different method and universe.
Market exposure explains part of the gap. At 1.09 beta, much of our 17.46% comes from the market, which the paper removes from its excess-return measure; that exposure also deepens our drawdown in selloffs. Our window includes the 2020 crash, used only for validation in the paper. Trading one expression concentrates the book more than a 1,000-factor tree model, consistent with our 28.33% volatility. We see two further likely contributors. The capitalization screen has not been verified as point-in-time and could contain survivorship. Our charge of $0.004 a share is probably cheaper for large caps than 20 basis points round trip, flattering our return. Position size may matter as well: our specification permitted up to 4.0x leverage, and any use of it could have increased both return and drawdown. Specified next-open fills also conflict with the recorded market-on-close execution. We cannot calculate our index-relative return and therefore cannot fully account for the gap.
Where implementation remains unclear
The paper specifies the model reward, all four validity horizons, DAPO clip ratios of 0.2 and 0.28, and the complete XGBoost settings. The US trading assumptions are less settled. The only cost schedule we found includes a 5 CNY minimum fee and 100 million CNY in starting capital; we did not find a separate US schedule. Nor did we find whether the WRDS constituent lists are point-in-time. The five seeds behind the ±0.0065 band for the 8B S&P 500 figure appear to change the downstream model. We found no reported variation across RL runs or generation draws. The 2019 cutoff applies to the prompts, while Qwen3's pretraining may extend beyond the 2021 to 2024 test period; we found no discussion of that possibility.
I would rebuild the US downstream stage with Ridge and style neutralization before treating the literature-conditioned RL as verified. Until a US run produces positive linear IC from these factors, I would size the 6.94% as a claim about a language-model-fed tree with 1,000 features.
Our backtest stops at 2024-07-01, and everything after that date is deliberately left untouched so the same strategy can be checked out of sample later.