Chen's ablation gives reinforcement-learning weight generation more credit than the sentiment feature named in the paper. Taking out RL policy optimization raises MSE from 0.0038 to 0.0052 and cuts Sharpe from 1.048 to 0.891. Removing the willingness module produces MSE of 0.0047, Sharpe of 0.942 and return of 0.177. Chen states this ordering directly. After removing RL strategy optimization, he writes, "performance degrades most significantly."

The baseline problem is acknowledged too: "Joint optimization under the unified PO objective function L is not conducted for these models." Chen argues that the components reinforce each other. The willingness module and RL strategy optimization, he writes, "jointly constitute the main sources of TransAWP's prediction accuracy improvement." Table 6 supports that reading. It also raises an awkward question: why do the title and abstract foreground willingness when the ablation ranks it second?

The architecture

TransAWP, Chen's name for his architecture, uses a Transformer encoder to produce portfolio weights directly. Its inputs come from five public Kaggle datasets, arranged as four input blocks, with daily coverage from 2018-01-01 to 2023-12-31. Two contain prices. Another provides both macro and asset-risk features, while two more supply sentiment extracted from financial news and social-media text. Monthly macro observations are carried forward to daily frequency. Sentiment records are assigned to the nearest trading day through merge_asof. Everything is Z-scored.

Three sentiment scalars make up the behavioral input: sentiment_score, polarity, emotion_score. The model embeds them in a 16-dimensional vector, tiles that vector across all T time steps and joins it to a 64-dimensional market feature block. Chen names the result investor adoption willingness.

The same vector also changes attention inside the encoder. Before the softmax, the model adds tanh(W_h H + W_w W_u + b) to the attention logits and scales the addition by a coefficient alpha. Group sentiment can therefore shift the time steps receiving attention.

A policy function generates the weights and is trained by policy gradient. Portfolio return drives the reward, with a quadratic risk penalty subtracted and a willingness-matching term w'W_u added. A VaR-style risk constraint and squared turnover penalty are layered onto that objective. The baselines use another construction. LSTM, GRU and the standard Transformer forecast next-day returns, then apply a Softmax to obtain weights. DRL-Portfolio and TransAWP produce weights directly.

TransAWP reports a 19.6% annualized return out of sample, without trading costs. Volatility is 18.7%, Sharpe is 1.048 with the risk-free rate set to zero, max drawdown is 21.1% and daily turnover is 0.261. LSTM has the weakest row at 14.2% / 21.5% / 0.660 / 28.4% / 0.318. The standard Transformer posts 16.7% / 19.8% / 0.843 / 24.6% / 0.289. DRL-Portfolio comes closest on return at 17.3%; its Sharpe of 0.844 is effectively tied with the standard Transformer's 0.843.

Hold the weight generator constant

Chen's qualification about objective function L determines how much confidence the portfolio comparison deserves. He repeats the warning: differences in the table "partly reflect the impact of different weight construction mechanisms, not just the differences in prediction accuracy." The ablation results allow a closer comparison.

One of the three variants is w/o RL. It retains the willingness embedding and modulated attention, while policy-gradient weight generation is replaced with predict-then-Softmax. Chen says this is "consistent with the weight generation methods of LSTM, GRU, and standard Transformer." As the paper describes them, w/o RL and the standard Transformer baseline therefore differ in the willingness embedding and modulated attention. Chen leaves unclear whether the risk-constraint term and squared turnover penalty remain in w/o RL, so the following gap may capture those terms as well.

The rows are close. The ablation table gives Sharpe of 0.891 on the test set, 2018 to 2023, versus 0.843 in the out-of-sample portfolio table. Return is 0.165 against 0.167. Max drawdown is 0.249 against 0.246.

With the weight generator held constant as far as the description permits, willingness contributes roughly 0.048 of Sharpe. Return and drawdown are marginally worse.

That arithmetic is mine and crosses two tables with slightly different labels. Table 5 says out-of-sample; Table 6 says test set over 2018 to 2023. The comparison is indicative. Its direction agrees with Chen's headline ranking, and the feature appears more valuable inside the RL loop. Removing willingness from the full model lowers Sharpe by 0.106, from 1.048 to 0.942. Across the two predict-then-Softmax rows, the gap is about 0.048, with 0.891 against 0.843. We found a similar pattern in a matched EVaR comparison where the headline estimator differed by 0.003 Sharpe once the portfolios were aligned.

The paper's clearest finding concerns the risk term. Once it is removed, prediction and return barely move: MSE is 0.0042, R2 is 0.919, return is 0.184 and Sharpe is 0.967 against 1.048. Max drawdown widens from 0.211 to 0.258. Chen separates tail control from fit cleanly here.

Missing implementation choices

The target variable remains uncertain. R2 exceeds 0.93 in every one of six years, Pearson correlation stays above 0.94 each year, and the predicted-to-actual variance ratio remains between 0.98 and 0.99. Combined with R2 above 0.93, such a tightly held ratio reads to us like a price-level target rather than a return target. Yet Chen describes the baselines and w/o RL as forecasting next-day asset returns. We did not find a statement identifying the series behind TransAWP's MSE of 0.0038. The accuracy section refers to market price and return rate sequences.

The held-out share is also hard to place on the calendar. Chen specifies a chronological 70/30 partition, using the first 70% in-sample. He says random partitioning and overlapping rolling windows were not used. Elsewhere, he describes five-fold cross-validation "under the same training set and validation set split (70%:30%)." Figs. 1 and 2 are then discussed as six annual observations, while the portfolio section calculates return sequences across a unified interval from 2018 to 2023.

The stated window spans 72 months. Thirty percent is about 21.6 months, implying a held-out period beginning around March 2022 and lasting roughly twenty-one months. The reader cannot determine whether 19.6% applies to those twenty-one months or to all six years.

Table 3 reports parameter ranges: 3-6 layers, 4-8 heads, alpha 0.2-0.5, lambda1 0.3-0.5, beta 0.1-0.3 and mapping dimension 64-256. Eleven rows use this format. We did not find the final configuration or an explanation of how it was selected. Alpha at 0.2-0.5 and depth at 3-6 layers matter most.

Frictionless results

Turnover of 0.261 measures the average absolute weight change between adjacent trading days, with the portfolio rebalanced daily. Charging 5bps one-way on traded notional comes to roughly 3.3% a year. Annualized return falls from 19.6% to about 16.3%, and Sharpe to 0.87. At 10bps, the corresponding figures are about 13.0% and 0.70.

The baselines turn over faster, including LSTM at 0.318 and DRL-Portfolio at 0.332, so the same charge hurts them more. The ranking remains, with DRL-Portfolio near 0.64 at 5bps. Chen explicitly excludes costs, liquidity and estimation error in the discussion. The ordering is the result. The 1.048 comes from a frictionless portfolio.

What does willingness capture?

Investor adoption willingness is Chen's label for three sentiment scalars derived from news and social-media text. In the limitations, he concedes that the variable "does not fully reflect individual-level behavioral dynamics". The definition makes the same boundary clear, saying the construct "avoids direct observation of individual real investment decision-making behaviors". No adoption decision is observed in the study, leaving the behavioral claim untested in either direction.

The reward function also contains an unresolved dimensional issue. Its willingness-matching term is presented as a dot product between an N-dimensional weight vector and the 16-dimensional embedding. The paper never explains how those dimensions are aligned.

We could not run this.

The universe is identified only as "stocks included in Data1". We did not find an asset count, ticker list or membership rule anywhere in the paper, leaving no way to map the result to a tradable portfolio. The willingness input also depends on a social-media public-opinion feed that we do not hold. Our text coverage starts around 2020 rather than 2018. Although all five source datasets are linked publicly on Kaggle, the paper says "the datasets used and/or analyzed during the current study are available from the corresponding author Haobo Chen on reasonable request via e-mail", and we did not find a code link.

The next test needs the asset list and selected hyperparameters, followed by one table comparing TransAWP with the standard Transformer when both are trained under the same policy-gradient objective. If the 0.106 Sharpe gap from the w/o Willingness ablation survived that comparison, the sentiment channel would be worth building.