Investors/traders track a range of sentiments (consumer, investor, analyst, forecaster, management), searching for indications of the next swing of the psychological pendulum that paces financial markets. Usually, they view sentiment as a contrarian indicator for market turns (bad means good — it’s darkest before the dawn). These blog entries relate to relationships between human sentiment and the stock market.
Considerable research investigates investor disagreement, measured with very different observable metrics (proxies), as an asset return predictor. Is this research coherent? In their June 2026 paper entitled “Measuring Investor Disagreement: Proxies, Pitfalls, and a Path Forward”, Christian Goulding, Campbell Harvey and Hrvoje Kurtović review the body of research on investor disagreement proxies, including those based on analyst forecasts, return volatility, trading and short interest, institutional holdings, option prices and machine learning methods. They organize these proxies into families and address the limitations of each as a standalone measure. They further address how to combine metrics into a composite variable. Based on the body of research on investor disagreement and data to compute the main investor disagreement proxies for U.S. common stocks priced over $5 during 1994 through 2025, they conclude that:Keep Reading
Can large language models (LLM) simulate market-scale, real-time business sentiment from the Chief Financial Officer (CFO) perspective? In their June 2026 paper entitled “CFOs Meet LLMs”, John Graham, Campbell Harvey and Manish Jha use an LLM (GPT-5.4) to simulate CFOs of specific firms. Simulations involve two prompts for each of 6,075 actual responses to the quarterly Duke-Federal Reserve CFO Survey during 2002 through 2025:
System prompt – establishes context, persona and analytical instructions, including instructions to collect from the web CFO interviews and analyst reports, price targets and consensus revenue estimates. Instructions include a temporal restriction to ensure that information collected was available before the matched survey response date.
User prompt – provides firm profile (industry, revenue, headcount and geographic footprint), respondent history (prior CFO optimism scores when available and the LLM’s own prior forecasts) and the key survey question: “Rate your optimism about the overall U.S. economy on a scale from 0–100, with 0 being the least optimistic and 100 being the most optimistic.” Again, a temporal restriction imposes an information cutoff date matching the actual survey date.
They run each pair of prompts three times a few seconds apart and average outputs to suppress LLM statistical variation. Finally they match every LLM average output to the response of the actual CFO at the same firm in the same quarter. Using past CFO Survey results and associated firm data spanning 2002 through 2025, they find that:Keep Reading
The quarterly CFO Survey asks chief financial officers, owner-operators, vice presidents and directors of finance, accountants, controllers, treasurers and others with financial decision-making roles in small to very large companies across all major industries to “rate optimism about the overall U.S. economy on a scale from 0 to 100.” Does the average economic sentiment of these financial experts predict U.S. stock market returns? To investigate, we relate quarterly sentiment averages and quarterly changes in these averages to quarterly S&P 500 Index (SP500) returns. Using the specified quarterly data during June 2002 through March 2026, we find that:
Are broad measures of public sociopolitical sentiment relevant to investors? Do they predict stock returns as indicators of exuberance and fear? To investigate, we relate S&P 500 Index return and 12-month trailing S&P 500 price-operating earnings ratio (P/E) to the percentage of respondents saying “yes” to the recurring Gallup polling question: “In general, are you satisfied or dissatisfied with the way things are going in the United States at this time?” Since individual polls span several days, we use S&P 500 Index levels for about the middle of the polling interval. To calculate market P/E, we use current S&P 500 Index level and most recently available quarterly aggregate operating earnings for that time. Using Gallup polling results, S&P 500 Index levels and 12-month trailing S&P 500 operating earnings as available during July 1990 (when polling frequency becomes about monthly) through March 2026, we find that:Keep Reading
Do exchange-traded funds (ETF) designed to exploit sentiment indicators beat the market? To investigate, we consider three such ETFs, all currently available, as follows:
VanEck Social Sentiment ETF (BUZZ) – invests in common stocks of U.S. companies with the most “positive insights” collected from online sources including social media, news articles, blog posts and other alternative datasets.
Relative Sentiment Tactical Allocation ETF (MOOD) – invests based on “relative sentiment” factors in other ETFs that hold equities, bonds, commodities, currencies and gold.
Stocksnips AI-Powered Sentiment US ALL Cap ETF (NEWZ) – invests in securities of U.S.-listed large, mid and small capitalization firms based on a proprietary, AI-derived News Media Sentiment Signal.
“Using the Money Anxiety Index for ETF Selection” examines whether the proprietary Money Anxiety Index (MAI) can select long and short portfolios of ETFs that beat the S&P 500 Index (ignoring dividends). Test outputs are 5-year, 3-year and 1-year cumulative returns. A deeper look at performance may be helpful. We extend the test period by 30 months and focus on the full period. We consider SPDR S&P 500 ETF Trust (SPY) and Invesco QQQ Trust (QQQ), which is much more like the MAI-selected ETFs, as benchmarks. We compute monthly total return statistics, along with compound annual growth rates (CAGR) and maximum drawdowns (MaxDD). Using beginning-of-month, dividend-adjusted prices for the 10 ETFs in the MAI portfolios and the two benchmarks from the beginning of May 2018 through the beginning of March 2026, we find that:Keep Reading
In response to our inquiry about Goldman Sachs Panic Index data, Grok responded that the data are proprietary and unavailable. However, Grok offered “several high-quality public proxies and near-replicas…built by traders and quants using only freely available data. These reconstructions correlate extremely closely (often 0.90–0.98) with the snippets Goldman has shown clients over the years.” For one of these proxies, the percentile rank of VIX within its trailing 2-year window (0-100 scale), Grok provided a Python script to generate an historical daily series. Is it predictive of U.S. stock market returns? To investigate, we run the script to generate daily Panic Index Proxy data from the end of 2015 through November 2025 and relate the series to contemporaneous daily S&P 500 Index (SP500) returns. Using these two series, we find that:Keep Reading
Can Grok extract a useful weekly U.S. stock market sentiment metric from posts on X? To investigate, we ask Grok to each week for two years aggregate weekly U.S. stock market sentiment looking for at least 50 posts per week (ending Saturdays) and weighting each post sentiment according to its audience engagement (influence). For example, the Grok Sentiment for 2025-11-29 encompasses posts from 2025-11-23 through 2025-11-29. We then relate the resulting aggregate sentiment values and change in these values to S&P 500 Index (SP500) returns from the first open after measurement (usually the Monday open) to the close before the next measurement (usually the Friday close). Using the specified weekly inputs we find that:Keep Reading
Do analysts/investors predictably and exploitably misinterpret tones of earnings calls? In their October 2025 paper entitled “Do Investors Get It Right? Reaction Bias to Earnings Calls”, Zhenzhen Fan and Fred Liu study interactions between textual information in earnings conference calls and analyst revisions to their next-quarter earnings forecasts. Specifically, they:
For each stock and each following analyst: (1) link call transcripts to the closest pre-call and post-call analyst forecasts, and (2) measure analyst reaction as the fraction of the pre-announcement forecast error corrected in the post-announcement revision.
Use Shapley values to assess the contribution of each model feature to predictive power.
Across stocks and earnings calls, use a random forest model to test whether earnings call transcripts predict direction and magnitude of reaction bias. Use seven rolling years of the sample for model training/validation and test predictive accuracy each next year.
Each month sort stocks into fifths (quintiles) based on predicted reaction bias and form a value-weighted, monthly rebalanced portfolio that is long stocks with the highest signal and short the stocks with the lowest. Keep a stock within its quintile until the end of the month of its next earnings announcement.
Using earnings conference call transcripts, analyst earnings forecasts, actual earnings and monthly stock return data during 2006 through 2023, they find that:
Business media and expert commentators sometimes cite the monthly University of Michigan Consumer Sentiment Index as an indicator of U.S. economic and stock market health, generally interpreting a jump (drop) in sentiment as good (bad) for future consumption and stocks. The release schedule for this indicator is mid-month for a preliminary reading on the current month and end-of-month for a final reading. Is this indicator predictive of U.S. stock market behavior in subsequent months? Using monthly final Consumer Sentiment Index data and monthly levels of the S&P 500 Index as available during January 1978 through September 2025, we find that:Keep Reading
Become a CXO Member
Gain access to hundreds of premium investing research articles and CXO's trading strategies
