Li offers six factors, four nested models and five hypotheses, without a single estimate a desk could assess. The paper's only table explains which factors belong in each model.
The abstract says the model "is evaluated through nested comparisons with FF3, sentiment-augmented FF4, and FF5 specifications". It lists pricing errors, the GRS test, Fama-MacBeth regressions and out-of-sample performance. Taken on its own, that wording presents a completed study. Later in the same abstract, however, Li describes "a falsifiable design". The conclusion says the empirical judgment "must therefore remain conditional on the data." Li acknowledges the objection a sceptic would raise first. The question worth pursuing is whether running the proposed tests would yield an answer a desk could trade.
Only partly. The architecture makes sense, but one comparison mixes distinct changes. And three events offer little power for the proposed herding and reversal tests.
Building the sentiment factor
Li starts from a familiar mechanism. A-shares have heavy retail participation and constraints on short sales. Noise-trader risk and limits to arbitrage may allow prominent narratives to pull prices away from cash flows for longer than in developed markets. The stocks most vulnerable to that pressure should have a separate return component.
Fama-French supplies the baseline. FF3 contains the market, SMB (size) and HML (value); FF5 adds RMW (profitability) and CMA (investment). Li proposes a sentiment factor to capture returns those five leave unexplained. Attention pushes up sentiment-sensitive stocks that are difficult to arbitrage, followed by a partial retreat. H5 predicts "partial mean reversion after extreme positive sentiment shocks". None of this has an estimate in the paper.
The sentiment measure combines three inputs. A monthly Baker-Wurgler-style index is stripped of macro content and lagged one period. Daily news tone comes from a finance language model, while daily abnormal Baidu search volume (ASVI) measures attention. Li combines the three in a weighted sum, using PCA weights or weights fixed ex ante. The sum is then regressed on same-day factor returns and macro controls, either recursively or over the estimation sample. SENT⊥, the residual, represents the non-fundamental component.
Li distinguishes that index from a factor. To build the sentiment-mimicking factor (SMF), the design neutralizes firms by size and industry, sorts them by estimated sensitivity to lagged SENT⊥, and takes a value-weighted high-minus-low position. CASE-FF6 adds SMF to the market, RMW and CMA, together with China-adapted size and value factors intended to reduce shell-value contamination among the smallest listed firms. The proposed universe comprises non-financial Shanghai and Shenzhen stocks from CSMAR or WIND. ST names and price-limit days receive separate treatment. We found no sample period, firm count or estimation window for the sentiment betas.
What sign should search volume carry?
News tone has a sign; search volume has none. Bad news can lift ASVI while pushing tone down, leaving the inputs partly offsetting each other. PCA assigns weights according to their in-sample co-movement.
Li invokes Barber and Odeon's attention-driven buying argument. Retail investors tend to buy stocks that attract their attention, and China's short constraints sharpen that asymmetry. This supports a positive ASVI weight. It also supports setting the weight before seeing the data, an option the paper permits only "in robustness tests".
The keyword list poses a tougher question. It includes "Nvidia", "HBM", "Rubin", "Feynman" and firm names. According to the paper's own source, Nvidia officially identified Feynman at GTC; that source is dated 19 March 2026. The terms listed are named after the 2026 events. Applying that same list to the factor's earlier history would embed a hindsight choice of keywords. Li does not explain how earlier years' lists would be chosen. Recursive orthogonalization, one of the two permitted approaches, is defensible.
Li calls SMF "tradable". We found no treatment of the portfolio's own turnover, trading costs or capacity. The introduction itself acknowledges institutional constraints on short-selling, and the low-sensitivity leg requires shorts. A paper portfolio can serve a pricing test. For someone taking the position, H1 requires a non-zero SMF mean and SMF unspanned by the other factors, and those properties matter only net of borrow.
The FF5 comparison changes more than sentiment
The four comparisons use FF3, FF3 plus SMF, FF5, and CASE-FF6. Under H2, CASE-FF6 should deliver smaller absolute alphas than FF5 for sentiment-sensitive test assets. Li requires Model B to beat Model A and Model D to beat Model C "after accounting for the cost of an additional regressor."
D makes more than that additional-regressor change. Alongside SMF, it introduces SMB^CH and HML^CH in place of SMB and HML. Li cites Liu, Stambaugh and Yuan precisely because constructing size and value for China substantially improves explanatory power. That replacement alone could account for a D-over-C win.
The FF3-to-FF3-plus-SMF comparison isolates SMF cleanly. It nevertheless uses the US-style size and value factors that Li argues suffer from shell-value contamination.
Li describes CASE-FF6 as adding SMF to a China-adapted Fama-French five-factor baseline. That baseline is absent from the four models being compared.
Comparing it with CASE-FF6 on the same test assets would assign the GRS and alpha differences to sentiment alone. Every required input is already in the design, so this comparison costs nothing extra.
Three Nvidia days in 2026
The event tests centre on CES (the paper's source is dated 5 January 2026), GTC (source dated 19 March 2026) and Huang's Seoul remarks on 7 June 2026. Li proposes a [-5,+10] window for attention pressure, narrowing it to [0,+2] for the immediate response. The framework's value lies in its ability to distinguish attention from cash-flow news.
Attention pressure predicts three things. ASVI should jump. Turnover and buying should cluster in stocks associated with memory, packaging and chip design, including those with weak Nvidia links. Prices should then partly reverse over the rest of the [-5,+10] window. Cash-flow news implies no reversal, with price moves concentrated among firms with genuine exposure.
Seoul makes the distinction harder. Huang said shortages spanning wafers and packaging through silicon photonics would last several years. A Chinese memory or packaging supplier could plausibly receive cash-flow news from those remarks. For that date, the entire test depends on separating firms by supply-chain exposure. The paper names that split without defining it.
Power remains limited. Three events tied to AI hardware amount to three draws of a common shock. Concept stocks in the same basket all receive the day-0 shock, so pooling them would make a cross-sectional t-statistic overstate precision. Li acknowledges that attributing an optical-stock selloff after GTC to forced liquidation requires "transaction-level or margin-financing evidence". The paper also says the events "should be used as test windows rather than as substitutes for empirical evidence". We agree with that standard. Three windows can meet it only as probes.
Why we could not backtest CASE-FF6
We could not run the test ourselves. Our data lacks Shanghai and Shenzhen histories for prices, fundamentals and trading status. Price-limit and suspension flags are therefore missing too. We also lack Baidu search history, one of the three sentiment inputs. Substituting US equities would remove both the short-sale constraint and the Nvidia-to-Chinese-concept-stock channel at the centre of the paper.
One result would change our assessment: SMF alpha surviving a China-adapted FF5 after costs, with the short leg confined to stocks actually available to borrow and keywords fixed before the sample begins. Until that number appears, CASE-FF6 remains a well-organized plan with one confounded comparison.