Tao's sentiment chain offers no measured trading edge

Tao's review gives a trader no evidence that news sentiment earns money after costs, and makes no claim to do so. Despite a title promising stock market prediction, Tao presents the paper as preparation for empirical research. It is a narrative literature review. The abstract describes it as systematic, yet supplies neither a search protocol nor inclusion criteria. There is no dataset or sample period, no universe or backtest. Every number below comes either from work Tao cites or from our adaptation. Those sources measure different things throughout.

Tao's starting argument is straightforward: conventional sentiment proxies arrive late because they depend on trades already made. Closed-end fund discounts, turnover, IPO first-day returns and margin balances all have that limitation. Surveys brought problems of their own. Drawing on Li Xinhua, Tao describes questionnaires as the main measurement method before 2000. China's CCTV Watch index (the BSI) asked 100 domestic securities investors about market direction, with institutional analysts overrepresented and results published after a delay. Its up-or-down answers could capture neither intensity nor disagreement. Tao places the machine-learning breakthroughs in 2010 and traces the text-sentiment approach to Tetlock (2007). Text classification offers a way around these delays: it can read news, forum posts and social media as they arrive, separating media tone from retail feeling.

The paper's second half sets out how sentiment might reach prices. It proposes three stages. First, biased media coverage shapes retail mood, an argument drawn from Mullainathan and Shleifer and from Tetlock. Noise traders convert that mood into orders, drawing on the De Long, Shleifer, Summers and Waldmann (1990) model and Black (1986). Herding amplifies the resulting move, with citations to Bikhchandani et al.'s information-cascade model (1992) and Banerjee's herding model (1992). In Tao's account, media coverage forms expectations, noise trading moves prices and herding increases volatility. The potential trading profit comes from retail-heavy order flow driving prices away from value. The paper then says, "stock prices usually return to fundamentals." It closes with two research gaps, three risks introduced by generative AI (information pollution, algorithm convergence and black-box models) and three governance proposals.

Who grades the three stages?

Tao joins separate literatures into a chain. None of the cited studies measures the full sequence, from media tone to retail tone, then noise orders and finally herding. The paper nevertheless assigns evidence grades to the individual stages. Media receives a relatively high rating; noise trading and traditional herding receive high ratings; AI-amplified herding receives a moderate rating. Tao gives no criteria for assigning them.

One concession deserves credit. Tao acknowledges that the AI-herding literature relies mainly on panel regressions and behavioral experiments, with large-sample, long-run time series evidence still needed. That qualification applies to stage three. The claim that all three stages make a closed loop receives no grade, although the conclusion treats it as a core insight.

For a trader, the reversal claim matters most. It comes without a horizon or magnitude. Tetlock's own result is the nearest evidence among the cited work: his 2007 paper found that high pessimism in a daily Wall Street Journal column predicted downward price pressure, followed by reversion over the next few trading days. Tao cites Tetlock for prediction alone. The finding describes a temporary price effect that subsequently reverses. Taken literally, it favors fading the move or holding for a day or two.

Faster scores, unmeasured forecasting gains

The review runs together two claims that need separate treatment. Text classification arrives sooner than a turnover proxy; that is undisputed. Whether the earlier reading forecasts returns beyond information already in prices remains a separate question. Tao offers only one sentence comparing text with trading-based proxies, attributed to Bollen et al.: "text-based emotion indicators had better predictive performance than traditional trading indicators." There is no metric attached.

Bollen et al. (2011) forecast DJIA direction relative to a baseline model. As we read the study, they did not compare their forecasts with trading-based sentiment proxies, and their result is difficult to replicate. The citations weaken the argument further. References [5] and [11] both point to that Bollen paper. Tao uses the second to support a different proposition, that emotion-driven noise trading causes short-term excess volatility. Predicting index direction from mood does not establish that claim.

The reference list also omits the standard sentiment benchmarks: the Baker and Wurgler index, the Loughran-McDonald finance dictionaries, and FinBERT. The review therefore cannot establish how much a text score adds over those alternatives.

Our US news-quintile book lost 23.40%

We made several substitutions to turn the idea into a trading test. Tao specifies no trading algorithm, leaving us to choose the classifier, portfolio rule and holding period. This is an adaptation of a hypothesis derived from the review. With no stock-forum or social-platform data available, we tested the news component alone. That excludes the media-to-retail transmission channel at the heart of the stage-one claim.

We also substituted US equities for the A-shares in Tao's examples. Retail participation and the strength of transmission may differ across those markets, so a US result says nothing about the A-share claims. Our earlier Chinese sentiment note found an effect in-sample only (CSI sentiment and CSI 300 tails). We had intended to use a 2019 FinBERT model for scoring. The run behind the figures below instead used a fixed finance lexicon, with negation handled within three tokens. These results cannot serve as a FinBERT baseline.

We constructed the portfolio as follows:

• We select the top 500 non-ADR US stocks by prior-year capitalization and refresh the universe annually.
• At each session, we score company news released after the previous close through the current close. We deduplicate by URL or text hash, then average scores for each stock.
• At the next open, we buy the top quintile and short the bottom quintile, holding each cohort for five sessions.
• Each leg receives 10% of equity. Gross exposure is capped at 4 times initial capital.

Our book lost 23.40% between January 2020 and July 2024. Tao reports no performance figures, leaving no paper result against which to compare it. The run produced a Sharpe of -1.26 and a Sortino of -1.40, with maximum drawdown of -29.72%. Annualized volatility was 4.14%; Calmar was -0.19.

It bled about 5% a year at low volatility.

This quick automated construction found no edge. The figures should first be read as evidence about our implementation. Its lexicon scores text less well than FinBERT. We chose the five-session holding period ourselves, and Tetlock's reversal over a few days suggests that choice may be too long. Names without news disappear from the daily selection, concentrating the book in heavily covered stocks.

We used neither momentum nor news-volume controls. Those exposures may account for part of the -23.40%, rather than tone itself. Before trading, the same window needs evaluation strictly in chronological order, with momentum and article counts controlled for and timestamps audited for leakage. Tao discusses none of these implementation decisions. They all remain the practitioner's responsibility.

Synthetic news reaches the feed first

Information pollution is the review's most useful practical warning. Generative models can manufacture false company announcements at scale, misleading retail investors and contaminating the "sentiment indicators" extracted through text mining. A news-scoring book faces that data-quality problem directly in its feed, before any market-level amplification occurs.

Algorithm convergence has a similarly concrete interpretation. Desks applying similar models to the same feed can end up with correlated, crowded signals. You can measure that correlation between your signal and vendor scores without assuming a market-wide effect.

The sourcing for amplification is tangled. Reference [6], attributed to a "National Research Network," first supports the claim that generative AI filters out false opinion. Later, as "Guoyan Network," it supports the claim that generative AI induces algorithmic herding. The former presents generative AI as cleaning the sentiment signal; the latter presents it as distorting the signal. The text attributes the "algorithm imitation effect" to "Uliari," yet reference [17] duplicates [16], a paper by Li Bingyu. None of the three risks has a magnitude attached.

Tao offers a reasonable reading list for anyone building a short-horizon A-share news signal, and weak grounds for expecting that signal to pay. A multi-day sentiment continuation that survives chronological evaluation, with momentum and news volume controlled for, would change that view.

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.

Before testing:
  Verify the archived classifier, tokenizer, labels and training/release dates.
  Freeze preprocessing and inference; perform no return-based fitting.
  The supplied code instead declares a fixed finance-lexicon substitute;
  treat that implementation as a deviation, not a compliant FinBERT baseline.

For each trading year y:
  Query screening year y-1 for STOCK securities with is_adr=false.
  Sort capitalization descending, then symbol_key ascending; retain 500.
  Freeze new-cohort membership for the year; preserve existing cohorts.
  Skip new cohorts if the required screen is unavailable.

At each actual exchange close t:
  Collect usable company news published in (previous close, close t].
  Require available article versions by cutoff wherever timestamps permit.
  Deduplicate causally per stock using normalized URL or text hash.
  Score each unique article; specified score = P_positive - P_negative.
  Average scores per stock; exclude stocks without eligible news.
  Sort by mean score ascending, then symbol_key ascending.
  Set k = floor(number eligible / 5).
  Skip formation if k=0 or all scores are identical.
  Select first k stocks short and last k stocks long, without sign filters.
  Do not schedule cohorts whose planned exit exceeds the test endpoint.

At the following opening event t+1:
  Process scheduled exits before entries using genuine adjusted opens.
  Budget 10% of contemporaneously available equity to each new cohort leg.
  Allocate each leg equally across k names; shares = allocation / adjusted open.
  Enforce aggregate absolute same-stock exposure <=10% of current equity.
  Enforce pre-netting gross exposure / initial capital <=4.
  Reduce or skip affected entries without reallocating unused budgets.
  Skip missing entry opens and known failed short locates without replacement.

Maintain separate signed lots for every cohort, including opposing lots.
Keep adjusted-share quantities fixed, except necessary corporate actions.
Exit at adjusted open t+6, five trading sessions after entry.
If an exit open is missing, retain exposure and postpone to the next genuine open.
Flag unresolved suspensions/delistings; never invent fills or settlements.
Value positions using adjusted closes; do not double-count distributions.
Apply the platform-recorded per-fill cost recalculation to reported metrics.