Sentiment Indicators

Do economic expectations of sophisticated investors relative to those of unsophisticated investors predict stock market returns? In the September 2020 revision of his paper entitled “Relative Sentiment and Machine Learning for Tactical Asset Allocation”, flagged by a subscriber, Raymond Micaletti investigates use of relative Sentix sentiment for tactical asset allocation. He each month constructs relative sentiment factors for regional U.S., Europe, Japan and Asia ex-Japan equity markets as differences in 6-month economic expectations between respective institutional and individual investors. He then applies machine learning algorithms to test 990 alternative strategies of relative sentiment for each region, augmented by both cross-validation and adjusted for data snooping. He tests usefulness of the most significant backtest results in two ways:

1. Translation of relative sentiment to equity allocations ranging from 0% to 100% for each equity market, with the non-equity allocation going to either bonds or cash. As benchmarks, he uses the average monthly equity allocation of relative sentiment strategies, with the balance allocated to bonds or cash, rebalanced monthly.
2. Ranking of regions by relative sentiment to predict which equity markets will be outperformers and underperformers next month.

Using monthly Sentix sentiment data as described, monthly returns for associated equity market indexes and spliced exchange-traded funds (ETF) and monthly returns for the Barclays US Aggregate Bond Index during August 2002 through September 2019 (with a 3-month gap in sentiment data during October 2002 through December 2002), he finds that: Keep Reading

In reaction to “Federal Reserve Holdings and the U.S. Stock Market”, a subscriber suggested analysis of market reactions to announcements (starts/ends) of major Federal Reserve System interventions, such as the series of quantitative easing (QE) initiatives. Reactions to such announcement should precede changes in actual holdings. To investigate, we look at cumulative returns of SPDR S&P 500 (SPY) and iShares Barclays 20+ Year Treasury Bond (TLT) during the 30 trading days after each of the following announcements:

• 11/25/08: QE-1 initiated
• 3/16/09: QE-1 expanded
• 3/31/10: QE-1 terminated
• 11/3/10: QE-2 initiated
• 6/29/12: QE-2 terminated
• 9/13/12: QE-3 initiated
• 12/12/12: QE-3 expanded
• 10/29/14: QE-3 terminated
• 3/23/20: “QE-4” initiated

Using daily dividend-adjusted prices for SPY and TLT spanning these dates, we find that: Keep Reading

“Verification Tests of the Smart Money Indicator” performs tests of ideas and setup features described in “Smart Money Indicator for Stocks vs. Bonds”. The Smart Money Indicator (SMI) is a complicated variable that exploits differences in futures and options positions in the S&P 500 Index, U.S. Treasury bonds and 10-year U.S. Treasury notes between institutional investors (smart money) and retail investors (dumb money) as published in Commodity Futures Trading Commission Commitments of Traders (COT) reports. Since findings for some variations in that test are attractive, we add two further robustness tests:

• Extend the sample period in “Verification Tests of the Smart Money Indicator” through the end of March 2020 to see how SMI performs during the 2019 coronavirus (COVID-19) stock market crash.
• Using the same setup as in “Verification Tests of the Smart Money Indicator”, test alternate definitions of commercial traders as smart money and non-commercial traders as dumb money (from legacy reports), rather than asset manager/institutional traders and non-reportables, respectively (from more granular disaggregated reports).

Using COT report data, dividend-adjusted SPDR S&P 500 (SPY) as a proxy for a stock market total return index, 3-month Treasury bill (T-bill) yield as return on cash (Cash) and dividend-adjusted iShares 20+ Year Treasury Bond (TLT) as a proxy for government bonds during 6/16/06 through 4/3/20, we find that:

Keep Reading

“Verification Tests of the Smart Money Indicator” reports performance results for a specific version of the Smart Money Indicator (SMI) stocks-bonds timing strategy, which exploits differences in futures and options positions in the S&P 500 Index, U.S. Treasury bonds and 10-year U.S. Treasury notes between institutional investors (smart money) and retail investors (dumb money). Do these sentiment-based results diversify those for the Simple Asset Class ETF Momentum Strategy (SACEMS) and the Simple Asset Class ETF Value Strategy (SACEVS)? To investigate, we look at correlations of annual returns between variations of SMI (no lag between signal and execution, 1-week lag and 2-week lag) and each of SACEMS equal-weighted (EW) Top 3 and SACEVS Best Value. We then look at average gross annual returns, standard deviations of annual returns and gross annual Sharpe ratios for the individual strategies and for equal-weighted, monthly rebalanced portfolios of the three strategies. Using gross annual returns for the strategies during 2008 through 2019, we find that: Keep Reading

A subscriber requested verification of findings in “Smart Money Indicator for Stocks vs. Bonds”, where the Smart Money Indicator (SMI) is a complicated variable that exploits differences in futures and options positions in the S&P 500 Index, U.S. Treasury bonds and 10-year U.S. Treasury notes between institutional investors (smart money) and retail investors (dumb money). To verify, we simplify somewhat the approach for calculating and testing SMI, as follows:

• Use a “modern” sample of weekly Traders in Financial Futures; Futures-and-Options Combined Reports from CFTC, starting in mid-June 2006 and extending into early February 2020.
• For each asset, take Asset Manager/Institutional positions as the smart money and Non-reporting positions as the dumb money.
• For each asset, calculate weekly net positions of smart money and dumb money as longs minus shorts. 
• For each asset, use a 52-week lookback interval to calculate weekly z-scores of smart and dumb money net positions (how unusual current net positions are). This interval should dampen any seasonality.
• For each asset, calculate weekly relative sentiment as the difference between smart money and dumb money z-scores.
• For each asset, use a 13-week lookback interval to calculate recent maximum/minimum relative sentiments between smart money and dumb money for all three inputs. The original study reports that short intervals work better than long ones, and 13 weeks is a quarterly earnings interval.
• Use a 13-week lookback interval to calculate final SMI as described in “Smart Money Indicator for Stocks vs. Bonds”.

We perform three kinds of tests to verify original study findings, using dividend-adjusted SPDR S&P 500 (SPY) as a proxy for a stock market total return index, 3-month Treasury bill (T-bill) yield as return on cash (Cash) and dividend-adjusted iShares 20+ Year Treasury Bond (TLT) as a proxy for government bonds. We calculate asset returns based on Friday closes (or Monday closes when Friday is a holiday) because source report releases are normally the Friday after the Tuesday report date, just before the stock market close. 

1. Calculate full sample correlations between weekly final SMI and both SPY and TLT total returns for lags of 0 to 13 weeks.
2. Calculate over the full sample average weekly SPY and TLT total returns by ranked tenth (decile) of SMI for each of the next three weeks after SMI ranking.
3. Test a market timing strategy that is in SPY (cash or TLT) when SMI is positive (zero or negative), with 0.1% (0.2%) switching frictions when the alternative asset is cash (TLT). We try execution at the same Friday close as report release date and for lags of one week (as in the original study) and two weeks. We focus on compound annual growth rate (CAGR) and maximum drawdown (MaxDD) as key performance metrics. Buying and holding SPY is the benchmark.

Using inputs as specified above for 6/16/06 through 2/7/20, we find that: Keep Reading

Do differences in expectations between institutional and individual investors in stocks and bonds, as quantified in weekly legacy Commitments of Traders (COT) reports, offer exploitable timing signals? In the February 2019 revision of his paper entitled “Want Smart Beta? Follow the Smart Money: Market and Factor Timing Using Relative Sentiment”, flagged by a subscriber, Raymond Micaletti tests a U.S. stock market-U.S. bond market timing strategy based on an indicator derived from aggregate equity and Treasuries positions of institutional investors (COT Commercials) relative to individual investors (COT Non-reportables). This Smart Money Indicator (SMI) has three relative sentiment components, each quantified weekly based on differences in z-scores between standalone institutional and individual net COT positions, with z-scores calculated over a specified lookback interval:

1. Maximum weekly relative sentiment for the S&P 500 Index over a second specified lookback interval.
2. Negative weekly minimum relative sentiment in the 30-Year U.S. Treasury bond over this second lookback interval.
3. Difference between weekly maximum relative sentiments in the 10-Year U.S. Treasury note and 30-year U.S. Treasury bond over this second lookback interval.

Final SMI is the sum of these components minus median SMI over the second specified lookback interval. He considers z-score calculation lookback intervals of 39, 52, 65, 78, 91 and 104 weeks and maximum/minimum relative sentiment lookback intervals of one to 13 weeks (78 lookback interval combinations). For baseline results, he splices futures-only COT data through March 14, 1995 with futures-and-options COT starting March 21, 1995. To account for changing COT reporting delays, he imposes a baseline one-week lag for using COT data in predictions. He focuses on the ability of SMI to predict the market factor, but also looks at its ability to enhance: (1) intrinsic (time series or absolute) market factor momentum; and, (2) returns for size, value, momentum, profitability, investment, long-term reversion, short-term reversal, low volatility and quality equity factors. Finally, he compares to several benchmarks the performance of an implementable strategy that invests in the broad U.S. stock market (U.S. Aggregate Bond Total Return Index) when a group of SMI substrategies “vote” positively (negatively). Using weekly legacy COT reports and daily returns for the specified factors/indexes during October 1992 through December 2017, he finds that: Keep Reading

Are there relationships between (1) the stock market outlook expressed by a U.S. equity mutual fund manager in semi-annual reports and (2) positioning and performance of that fund? In his October 2019 preliminary paper entitled “Are Professional Investors Prone to Behavioral Biases? Evidence from Mutual Fund Managers”, Mehran Azimi examines these relationships. Specifically, for each such U.S. equity mutual fund semi-annual report, he:

1. Uses a word list to identify parts of fund reports that may contain stock market outlooks.
2. Applies machine learning to isolate sentences most likely to present outlooks.
3. Manually reads and rates these sentences as bearish, neutral or bullish.
4. Computes fund manager “Belief” as number of bullish sentences minus number of bearish sentences divided by the total number of sentences isolated. Positive (negative) Belief indicates a net bullish (bearish) outlook.

He then employs regressions to relate fund manager Belief to fund last-year return, asset allocation, portfolio risk and next-year 4-factor (adjusting for market, size, book-to-market and momentum) alpha. Using 40,731 semi-annual reports for U.S. equity mutual funds and associated fund characteristics, holdings and returns during February 2006 through December 2018, he finds that:

Keep Reading

Can investors apply deep learning software to expose obscure but useful management sentiment in firm SEC Form 10-K filings? In their July 2019 paper entitled “Is Positive Sentiment in Corporate Annual Reports Informative? Evidence from Deep Learning”, Mehran Azimi and Anup Agrawal apply deep learning to detect positive and negative sentiments at the sentence level in 10-Ks. They train their model using 8,000 manually evaluated sentences randomly selected from 10-Ks. They then use the trained model to assign sentiments to all sentences in each 10-K. Their overall measure of negative (positive) sentiment is number of negative (positive) sentences divided by the total number of sentences in the 10-K. They assess impact of 10-K sentiment on stock performance based on 4-factor (market, size, book-to-market, momentum) alpha during short intervals after 10-K filing. Using 10-K filings for non-utility and non-financial U.S. public firms with at least 200 words, associated daily stock prices/trading volumes and daily 4-factor alphas during January 1994 through December 2017, they find that: Keep Reading

Do sentiment indexes usefully predict U.S. stock market returns? In his May 2018 doctoral thesis entitled “Forecasting Market Direction with Sentiment Indices”, flagged by a subscriber, David Mascio tests whether the following five sentiment indexes predict next-month S&P 500 Index performance:

1. Investor Sentiment – the Baker-Wurgler Index, which combines six sentiment proxies.
2. Improved Investor Sentiment – a modification of the Baker-Wurgler Index that suppresses noise among input sentiment proxies.
3. Current Business Conditions – the ADS Index of the Philadelphia Federal Reserve Bank, which combines six economic variables measured quarterly, monthly and weekly to develop an outlook for the overall economy.
4. Credit Spread – an index based on the difference in price between between U.S. corporate bonds and U.S. Treasury instruments with matched cash flows. (See “Credit Spread as an Asset Return Predictor” for a simplified approach.)
5. Financial Uncertainty – an index that combines forecasting errors for large sets of economic and financial variables to assess overall economic/financial uncertainty.

He also tests two combinations of these indexes, a multivariate regression including all sentiment indexes and a LASSO approach. He each month for each index/combination predicts next-month S&P 500 Index return based on a rolling historical regression of 120 months. He tests predictive power by holding (shorting) the S&P 500 Index when the prediction is for the market to go up (down). In his assessment, he considers: frequency of correctly predicting up and down movements; effectiveness in predicting market crashes; and, significance of predictions. Using monthly data for the five sentiment indexes and S&P 500 Index returns during January 1973 through April 2014, he finds that: Keep Reading

How does news flow interact with short-term stock market return? In their April 2019 paper entitled “Forecasting the Equity Premium: Mind the News!”, Philipp Adämmer and Rainer Schüssler test the ability of a machine learning algorithm, the correlated topic model (CTM), to predict the monthly U.S. equity premium based on information in news articles. Their news inputs consist of about 700,000 articles from the New York Times and the Washington Post during June 1980 through December 2018, with early data used for learning and model calibration and data since January 1999 used for out-of-sample testing. They measure the U.S. stock market equity premium as S&P 500 Index return minus the risk-free rate. Specifically, they each month:

1. Update news time series arbitrarily segmented into 100 topics (with robustness checks for 75, 125 and 150 topics).
2. Execute a linear regression to predict the equity premium for each of the 100 topical news flows.
3. Calculate an average prediction across the 100 regressions.
4. Update a model (CTM^Sw) that switches between the best individual topic prediction and the average of 100 predictions, combining the flexibility of model selection with the robustness of model averaging.

They use the inception-to-date (expanding window) average historical equity premium as a benchmark. They include mean-variance optimal portfolio tests that each month allocate to the stock market and the risk-free rate based on either the news model or the historical average equity premium prediction, with the equity return variance computed from either 21-day rolling windows of daily returns or an expanding window of monthly returns. They constrain the equity allocation for this portfolio between 50% short and 150% long, with 0.5% trading frictions. Using the specified news inputs and monthly excess return for the S&P 500 Index during June 1980 through December 2018, they find that:

Keep Reading