Sentiment Indicators

“Mark Hulbert’s NASDAQ Newsletter Sentiment Index” reviews the usefulness of the Hulbert Stock Newsletter Sentiment Index (HSNSI), which “reflects the average recommended stock market exposure among a subset of short-term market timers tracked by the Hulbert Financial Digest.” Mark Hulbert presents HSNSI as a contrarian signal for future stock returns; when HSNSI is high (low), he views the outlook for stocks as materially bearish (bullish). In recent years, he has shifted emphasis in his MarketWatch columns from HSNSI to the Hulbert Nasdaq Newsletter Sentiment Index (HNNSI), stating that: “Since the Nasdaq responds especially quickly to changes in investor mood, and because those timers are themselves quick to shift their recommended exposure levels, the HNNSI is the Hulbert Financial Digest’s most sensitive barometer of investor sentiment.” Is HNNSI useful? Using a small sample of 38 values of HNNSI over the period April 2010 through September 2015 (generated by searching MarketWatch.com for “HNNSI”) and contemporaneous daily closes of the S&P 500 Index, we find that: Keep Reading

Are readily available crowdsourced firm earnings estimates and stock sentiment measurements exploitable? In the September 2015 revision of their paper entitled “Tweet Sentiments and Crowd-Sourced Earnings Estimates as Valuable Sources of Information Around Earnings Releases”, Jim Kyung-Soo Liew,  Shenghan Guo and Tongli Zhang investigate whether earnings estimates from Estimize and sentiment measurements from iSentium usefully predict stock behavior after earnings announcements. Estimize aggregates inputs from students, independent researchers, private investors, sell-side professionals and buy-side analysts to generate earnings estimates. iSentium derives sentiment scores (ranging from -30 to +30) from real-time natural language processing of Twitter texts about stocks, market indexes and exchange-traded funds. The authors relate pre-announcement earnings estimates and sentiment to post-earnings announcement stock returns. Using Estimize and iSentium data as available, Wall Street consensus earnings estimates, actual firm quarterly earnings and associated stock returns for 16,840 earnings announcements during November 2011 through December 2014, they find that: Keep Reading

Does firm news reliably interact with stock return anomalies? In their July 2015 paper entitled “Anomalies and News”, Joseph Engelberg, David McLean and Jeffrey Pontiff compare anomaly returns on days with and without firm-specific news releases. They consider 97 anomalies published in 80 academic papers. For some analyses, they segregate these anomalies into four categories: (1) firm event-related (such as stock issuance); (2) market (such as momentum); (3) valuation (such as earnings-price ratio); and, (4) fundamental (such as acruals). They measure each anomaly using the extreme fifths (quintiles) of monthly stock sorts to specify a long side and short side. They calculate returns in three-day intervals around news days. Using stock and firm data required to construct anomaly portfolios, 489,996 earnings announcements and 6,223,007 Dow Jones news items during 1979 through 2013, they find that: Keep Reading

Are stock return anomalies strongest when investor sentiment is highest or liquidity lowest? In the January 2015 draft version of his paper entitled “What Explains the Dynamics of 100 Anomalies?”, Heiko Jacobs  addresses these questions. He first identifies, categorizes and replicates 100 well-known or recently discovered long-short stock return anomalies related to: violations of the law of one price, momentum, technical analysis, short-term and long-term reversal, calendar effects, lead-lag effects among economically linked firms, pairs trading, beta, financial distress, skewness, differences of opinion, industry effects, fundamental analysis, net stock issuance, capital investment and firm growth, innovation, accruals, dividend payments and earnings surprises. He measures the gross magnitude and direction of these anomalies via long-short extreme decile (stocks in top and bottom tenths as ranked by a specific variable) portfolios. He then examines how gross three-factor (market, size, book-to-market) alphas for these anomalies vary with:

• The Baker and Wurgler market-level investor sentiment index, adjusted for economic conditions.
• Six measures of limits to arbitrage (cost of anomaly exploitation): VIX, average idiosyncratic volatility, TED spread, Moody’s credit spread, average bid-ask spread and market illiquidity.

Using monthly data as available for a broad sample of U.S. stocks, excluding those that are relatively small and illiquid, as available during August 1965 through December 2011 (many tests start much later and end January 2011), he finds that: Keep Reading

A reader suggested looking at Rydex asset ratios as stock market sentiment indicators. The reasoning for these indicators is that a high (low) ratio of assets in bullish funds to assets in bearish funds indicates an overbought (oversold) market. Are these indicators useful? The most timing-intensive traders arguably use leveraged funds, suggesting that a bull-bear asset ratio for such funds may be especially informative and timely. We therefore use the ratio of daily closing asset level for the S&P 500 2x Strategy – H Class (RYTNX) mutual fund to daily closing asset level of the Inverse S&P 500 2x Strategy – H Class (RYTPX) mutual fund (Rydex 2X/-2X). Using daily asset levels for these funds from inception on 5/19/00 through January 2015, along with contemporaneous daily opens of the S&P 500 index (since fund assets are available only after the close), we find that: Keep Reading

Does interest in, or concern about, financial markets as expressed in Internet searches predict stock market behavior? In the December 2014 revision of their paper entitled “Can We Predict the Financial Markets Based on Google’s Search Queries?”, Marcelo Perlin, Joao Caldeira, Andre Santos and Martin Pontuschka investigate whether changes in Google search frequency for finance-related words predict changes in stock market index level, volatility and trading volume in four English speaking countries (U.S., UK, Australia and Canada). They select 15 relevant search words/terms by measuring the frequency of appearance in four finance textbooks of a large number of candidates from an online financial dictionary. They then use Google Trends to construct time series of relative search frequency (on a scale of 0 to 100) for the selected words/terms in each of the four countries and relate these series to respective country stock market behaviors. Finally, they test a timing strategy that is each week long or short an index depending on level of local Google Trends search activity. Using the search activity time series and daily levels and constituent trading volumes for major stock market indexes in the four countries (aggregated weekly) during January 2005 through December 2013, they find that: Keep Reading

Do expected investment returns as predicted by experts in surveys reliably predict actual future returns? In the October 2014 version of their preliminary paper entitled “Survey Expectations of Returns and Asset Pricing Puzzles”, Ralph Koijen, Maik Schmeling and Evert Vrugt compare survey-based expected returns to actual future returns for three major asset classes encompassing: 13 country equity market indexes; 19 currencies (versus the U.S. dollar); and, 10-year government bonds in 10 countries. They measure actual asset returns in U.S. dollars based on futures prices for equities and bonds (actual or synthetic) and forward returns for currencies. Survey-based expected returns derive from the quarterly World Economic Survey of experts, which solicits six-month expectations (“higher” or “about the same” or “lower”) for local equity prices, currency value versus the U.S. dollar and long-term government bond yield. The currency survey series commences the first quarter of 1989, while the equity and bond series commence the second quarter of 1998. They test the accuracy of survey expectations in two ways:

1. Cross-sectional hedge portfolios that are each month long (short) the rank-weighted assets with the highest (lowest) survey expectations.
2. Time series portfolios that are each month long (short) each asset depending on whether respective survey expectations indicate a positive (negative) return.

Analyses include testing of different lags between survey month and actual future return measurement, noting that a reliably executable strategy requires a lag of at least three months. Using quarterly survey response data and monthly futures/forward returns for the specified assets as available through September 2012, they find that: Keep Reading

Do hedge fund managers who use technical analysis beat those who do not? In their May 2014 paper entitled “Sentiment and the Effectiveness of Technical Analysis: Evidence from the Hedge Fund Industry”, David Smith, Na Wang, Ying Wang and Edward Zychowicz examine the relative performance of users and non-users of technical analysis among hedge fund managers in different sentiment environments. They hypothesize that short-selling constraints prevent market correction of mispricings when sentiment is high (overly optimistic), but not when sentiment is low (overly pessimistic). Discovery of mispricings via technical analysis may therefore be more effective when sentiment is high. To test their hypothesis, they compare the performance of hedge funds that report using technical analysis to that of hedge funds that do not, with focus on the state of market sentiment. They define the market sentiment state as high or low depending on whether the monthly Baker-Wurgler market sentiment measure is above or below its full-sample median. Using end-of-period status on use/non-use of technical analysis and monthly returns for 3,290 live and 1,845 dead funds from the Lipper TASS hedge fund database and monthly market sentiment data during January 1994 through December 2010, they find that: Keep Reading

Research (see “Asset Growth Rate as a Return Indicator” and “Asset Growth a Bad Sign for Stocks Everywhere?”) indicates that stocks of firms with high asset growth rates tend subsequently to underperform the market. Does this finding translate to the overall stock market? In the April 2014 version of his paper entitled “Asset Growth and Stock Market Returns: a Time-Series Analysis”, Quan Wen examines whether the asset growth anomaly observed at the firm level applies in aggregate to the U.S. stock market. He also investigates whether any aggregate effect is predominantly behavioral or risk-based. He estimates aggregate growth rate quarterly as the market capitalization-weighted sum of firm-level percentage changes in book value of total assets. To ensure all asset data is known to investors, he relates asset growth rate to returns two quarters later. Using quarterly U.S. stock market excess returns (relative to the risk-free rate), asset growth rates for listed U.S. firms that employ calendar year accounting, analyst forecasts/revisions, stock returns around earnings announcements, and data required for comparison of asset growth with other U.S. stock market indicators during 1972 through 2011, he finds that: Keep Reading

Can traders extract an edge from sentiments expressed about stocks via public Internet message boards? In the March 2014 draft of their paper entitled “Investor Sentiment from Internet Message Postings and Predictability of Stock Returns”, Soon-Ho Kim and Dongcheol Kim investigate whether combined investor sentiment as expressed in Yahoo!Finance stock message boards predicts returns, volatilities and trading volumes of associated stocks. They measure sentiment of posters in two ways: (1) standardized sentiments that some posters explicitly include in their messages (Strong Buy, Buy, Hold, Sell or Strong Sell); and, (2) sentiment from all posts inferred by a learning algorithm trained on the subset with explicit standardized sentiments. They group standardized sentiments as buy (Strong Buy or Buy) and sell (Strong Sell or Sell). They ignore the Hold sentiment. They combine posted sentiments in two ways using either the normalized number of buys minus number of sells or the logarithm of an adjusted ratio of buys to sells. When an author posts multiple messages for the same stock within a measurement interval, they use only the most recent message. They analyze predictive power of combined sentiment at daily, weekly and monthly horizons. Using sentiments from more than 32 million messages for 91 stocks posted on Yahoo!Finance message boards by over a half million authors, and contemporaneous daily prices and trading volumes for these stocks, during January 2005 through December 2010, they find that: Keep Reading