Public Sentiment and Stock Prices

Loading dated sentiment evidence and prices.

Findings in these examples

These are selected, documented cases. Returns below are calculated from the market-price history used by this site, starting at the next trading session's close. The controls above show each path and source; figures below use the November 17, 2026 snapshot, whose latest close is November 13.

  • Bubble warnings can precede gains or losses. After Grantham's January 2021 warning, QQQ gained 25.0% over 252 sessions. After his January 2022 warning, it lost 18.0% over the same horizon. These are two related observations from one investor, with different entry dates—not independent evidence for a trading rule. Neither warning supplied an exact crash date.
  • A recovery can follow further losses. After Fundstrat's July 22, 2022 bottom call, SPY was 5.4% lower after 63 sessions and reached a 9.9% closing loss from entry. At 252 sessions it was 15.2% higher. A positive eventual return would conceal the intervening loss. A loss from entry alone does not establish whether a previously named market low was breached.
  • Strong pessimism is not a precise entry signal. AAII bearishness reached 59.4% in April 2022 and 60.9% in September. Over the next 63 sessions, SPY's worst closing losses from entry were 11.2% and 3.1%, respectively. Similar survey extremes had different downside paths.
  • Bullish company views can have different outcomes. Following Pershing Square's January 2022 Netflix purchase announcement, Netflix lost 48.4% over 63 sessions. Following Altimeter's October 2022 Meta letter, Meta gained 109.7% over 252 sessions, after falling 35.3% below the entry close. The letter was conditional on business changes; these price paths do not measure the causal effect of either publication.
  • Concern about AI does not specify an equity forecast. After Pew's February 2023 publication, when 38% were more concerned than excited about AI, Nvidia gained 44.0% over 63 sessions. After the August publication, when that share was 52%, it lost 2.0% over 63 sessions but gained 141.0% over 252 sessions. The surveys measured attitudes toward AI in daily life, not expectations for Nvidia's shares.

The May 2026 Janus Henderson survey directly addresses bubble concern: 67% of the sampled affluent investors worried about a bubble or correction over the next year, while 61% expected AI to improve returns over five years. Those beliefs can coexist. The percentages describe respondents, not the probability of a collapse. As of November 17, 2026, neither stated horizon had elapsed.

What the research supports

Seven studies and their limits
  • Baker and Wurgler (2006): sentiment predicts differences in subsequent returns across stock types, especially difficult-to-value and hard-to-arbitrage firms. This is a conditional cross-sectional result, not a universal market-timing rule.
  • Tetlock (2007): pessimistic financial coverage precedes short-run price pressure and subsequent reversal. News tone is a particular media sample, not a representative public poll.
  • Antweiler and Frank (2004): stock-message-board activity helps predict volatility; its return effect is economically small. Statistical significance alone does not establish a useful trading return.
  • Greenwood and Shleifer (2014): surveyed investor expectations rise with past returns and market levels, while moving against model-based expected returns. Optimism may be a response to a rally. Model-based expected returns are not realized future returns.
  • Chen, De, Hu and Hwang (2014): Seeking Alpha article and comment tone predicts returns and earnings surprises in their sample. Some opinions convey information about fundamentals, so automatically taking the opposite side can discard useful information.
  • Barber and Odean (2008): attention affects individual investors' buying choices. A rise in search activity or mentions does not, by itself, measure bullishness or bearishness.
  • Loughran and McDonald (2011): generic negative-word dictionaries often misclassify financial language. The evidence here preserves the actual question or claim instead of assigning an opaque universal sentiment score.

Measurement

Dates, returns, proxies, and selection

Public availability. Each marker uses the publication date of the linked source. Survey fieldwork is recorded separately. The entry is the first available benchmark trading session strictly after publication, at its closing price. This deliberately excludes the publication-day reaction and the next session's intraday move. Historical fieldwork is never backdated into a tradable signal.

Fixed horizons. The 5, 21, 63, 126, and 252-session windows count trading sessions after entry. Return is P(h) / P(0) − 1. Excess return is the asset return minus SPY's return over the same dates, in percentage points; it is not risk-adjusted alpha. Prices are split-adjusted daily closes and exclude dividends, trading costs, and taxes. Missing prices are not interpolated; unfinished horizons remain pending.

Path risk. Worst loss is the lowest closing return relative to entry, including zero at entry. Maximum drawdown is the largest closing decline from a preceding peak within the window. Peak gain is the largest closing gain from entry. Daily closes can miss larger intraday moves. The preceding 21-session return ends on the last session on or before publication and helps show whether sentiment followed an existing move.

Instrument mapping. Nvidia is a single-company AI exposure; SMH represents semiconductor stocks; BOTZ is a robotics and AI thematic fund; QQQ represents a broader Nasdaq-100 exposure. They have different holdings and risks. These are comparison proxies, not interchangeable AI indexes. Company-specific observations map to the named stock. SPY supplies a broad U.S. equity benchmark; when SPY itself is selected, a separate comparison series is omitted.

Evidence classes. Public surveys, investor surveys, and individual investment commentary stay separate. AAII respondents are self-selected members; other surveys retain their own sampling descriptions. Concern about technology, product reputation, expensive valuation, and an explicit prediction of falling prices are different claims. The same person can expect both social harm and shareholder profits. Separate survey questions can have overlapping respondents; their percentages are not subtracted into a synthetic sentiment balance.

Selection. These 23 initial records were chosen for documented sources, differing views, and useful comparisons. They are not a random sample or a complete historical feed. Repeated issuers, overlapping windows, and related AI exposures are dependent. The collection supports case comparisons, not an estimated probability that a bubble warning, bottom call, or poll predicts the market. Earnings, rates, valuation changes, and other news can explain subsequent prices.

Tracking. Prices refresh from the site's market-data endpoint when the page opens; an explicitly dated saved snapshot remains available if the request fails or returns shorter backup coverage. Sentiment evidence is reviewed weekly, has its own review date, and changes only when another dated source is verified. Older observations retain their publication dates and original horizons. A prospective test would need a predefined source universe, complete inclusion, frozen classification rules, and evaluation on later observations before claiming a predictive effect.