Equity Premium

Referring to “Substitute QQQ for SPY in SACEVS and SACEMS?” and “Conditionally Substitute SSO for SPY in SACEVS and SACEMS?”, a subscriber requested a horse race for boosting the performance of the Simple Asset Class ETF Value Strategy (SACEVS) and the Simple Asset Class ETF Momentum Strategy (SACEMS), and thereby the Combined Value-Momentum Strategy (SACEVS-SACEMS), based on substituting:

1. ProShares Ultra S&P500 (SSO) for SPDR S&P 500 ETF Trust (SPY) in portfolio holdings, but not in SACEMS asset ranking calculations.
2. Invesco QQQ Trust (QQQ) for SPY in both portfolio holdings and SACEMS asset ranking calculations.

In conducting the horse race, we focus on gross compound annual growth rate (CAGR), maximum drawdown (MaxDD) and gross annual Sharpe ratio as key performance metrics. In Sharpe ratio calculations, we employ the average monthly yield on 3-month U.S. Treasury bills during a year as the risk-free rate for that year. Using monthly total (dividend-adjusted) returns for SACEVS assets, SACEMS assets, SSO and QQQ as available through February 2024, we find that: Keep Reading

Subscribers asked whether substituting Invesco QQQ Trust (QQQ) for SPDR S&P 500 (SPY) in the Simple Asset Class ETF Value Strategy (SACEVS) and the Simple Asset Class ETF Momentum Strategy (SACEMS) improves outcomes. To investigate, we substitute monthly QQQ dividend-adjusted returns for SPY dividend-adjusted returns in the two model strategies. We then compare the modified performance with the original baseline performance, including: gross compound annual growth rates (CAGR) at various horizons, average gross annual returns, standard deviations of gross annual returns, gross annual Sharpe ratios and maximum drawdowns (MaxDD) based on monthly measurements. In Sharpe ratio calculations, we employ the average monthly yield on 3-month U.S. Treasury bills during a year as the risk-free rate for that year. Using the specified methodology and data to generate SACEVS monthly returns starting August 2002 and SACEMS monthly returns starting July 2006, all through January 2024, we find that:

Keep Reading

Is ChatGPT useful for predicting stock market returns based on financial news headlines? In the December 2023 version of their paper entitled “ChatGPT, Stock Market Predictability and Links to the Macroeconomy”, Jian Chen, Guohao Tang, Guofu Zhou and Wu Zhu investigate whether ChatGPT 3.5 can predict U.S. stock market (S&P 500 Index) returns based on Wall Street Journal front-page news headlines/alerts. The instruction they give ChatGPT 3.5 is:

“Forget all previous instructions. You are now a financial expert giving investment advice. I’ll give you a news headline, and you need to answer whether this headline suggests the U.S. stock prices are GOING UP or GOING DOWN. Please choose only one option from GOING UP, GOING DOWN, UNKNOWN, and do not provide any additional responses.”

They first compute monthly ratios of good news to total news (NR^G) and bad news to total news (NR^B) and then relate these ratios to S&P 500 Index excess returns over the next 1, 3, 6, 9 or 12 months. They compare the ability of ChatGPT to predict returns to that of traditional human interpretation and to those of BERT and RoBERTa as ChatGPT alternatives. Using daily Wall Street Journal front-page news headlines/alerts and monthly S&P 500 Index excess returns during January 1996 through December 2022, they find that:

Keep Reading

Can enhanced machine learning models accurately time popular equity factors? In their January 2024 paper entitled “Multi-Factor Timing with Deep Learning”, Paul Cotturo, Fred Liu and Robert Proner explore equity factor timing via a multi-task neural network model (MT) to capture the commonalities across factors and a dynamic multi-task neural network model (DMT) to extract financial and macroeconomic states. They attempt to time six well-known factors: (1) excess market return, size, value, profitability, investment and momentum. They employ 272 model inputs (123 macroeconomic and 149 financial) to predict each month:

1. The sign of next-month return for each factor.
2. The return for an equal-weighted portfolio that holds the factors (the risk-free asset) for factors with positive (negative) predicted returns.

The compare performances of MT and DMT with those of seven simpler off-the-shelf machine learning models: logistic regression (LR), penalized logistic regression (EN), random forest (RF), extremely randomized trees (XRF), gradient boosted trees (GBT), support vector machine (SVM) and feed-forward neural network (NN). For all models, they use the first 20 years of their sample period for training, the next five years for validation and the remaining years for out-of-sample testing. Their benchmark is an equal-weighted portfolio of all six factors. Using monthly data for the 272 model inputs and monthly returns for the six factors during January 1965 through December 2021, with out-of-sample testing starting January 1990, they find that: Keep Reading

Can machine learning models pick stocks that unequivocally generate alpha out-of-sample? In their November 2023 paper entitled “The Expected Returns on Machine-Learning Strategies”, Vitor Azevedo, Christopher Hoegner and Mihail Velikov assess expected net returns and alphas of machine learning-based anomaly trading strategies. They use nine machine learning models to predict next-month stock returns based on inputs for up to 320 published anomalies, added to the mix according to respective publication dates:

• Ordinary Least Squares with Huber Loss Function (OLS-HUBER).
• Elastic Net (ENET).
• Feedforward Neural Network (FFNN) with two to five hidden layers (FFNN2, FFNN3, FFNN4 and FFNN5).
• Long Short-Term Memory (LSTM) with one or two hidden layers (LSTM1 and LSTM2).
• ENSEMBLE, a combination of all other models.

They train the models using an expanding window, with the last seven years reserved for six years of validation and one year of out-of-sample-testing. During the test year, they each month reform a portfolio that is long (short) the value-weighted tenth, or decile, of stocks with the highest (lowest) predicted next-month returns. They then calculate actual next-month gross returns and 6-factor (market, size, value, profitability, investment and momentum) alphas during the test year. To calculate net returns and alphas, they multiply trading frictions estimated from historical bid-ask spreads times monthly portfolio turnovers. Using returns and firm characteristics for a broad sample of U.S. common stocks having data covering at least 20% of the 320 anomalies during March 1957 through December 2021, with out-of-sample tests starting January 2005, they find that:

Keep Reading

Can explainable artificial intelligence (AI) bridge the gap between complex machine learning predictions and economically meaningful interpretations? In their December 2023 paper entitled “Empirical Asset Pricing Using Explainable Artificial Intelligence”, Umit Demirbaga and Yue Xu apply explainable artificial intelligence to extract the drivers of stock return predictions made by four machine learning models: XGBoost, decision tree, K-nearest neighbors and neural networks. They use 209 firm/stock-level characteristics and stock returns, all measured monthly, as machine learning inputs. They use 70% of their data for model training, 15% for validation and 15% for out-of-sample testing. They consider two explainable AI methods:

1. Local Interpretable Model-agnostic Explanations (LIME) – explains model predictions by approximating the complex model locally with a simpler, more interpretable model.
2. SHapley Additive exPlanations (SHAP) – uses game theory to determine which stock-level characteristics are most important for predicting returns.

They present a variety of visualizations to help investors understand explainable AI outputs. Using monthly data as described for all listed U.S stocks during March 1957 through December 2022, they find that:

Keep Reading

Can causal discovery algorithms (which look beyond simple statistical association, and instead consider all available data and allow for lead-lag relationships) make economically meaningful contributions to equity investing? In their December 2023 paper entitled “Causal Network Representations in Factor Investing”, Clint Howard, Harald Lohre and Sebastiaan Mudde assess the economic value of a representative score-based causal discovery algorithm via causal network representations of S&P 500 stocks for three investment applications:

1. Generate causality-based peer groups (e.g., to account for characteristic concentrations) to neutralize potentially confounding effects in long-short equity strategies across a variety of firm/stock characteristics.
2. Create a centrality factor represented by returns to a portfolio that is each month long (short) peripheral (central) stocks.
3. Devise a monthly network topology density market timing indicator.

Using daily and monthly data for S&P 500 stocks and monthly returns for widely used equity factors during January 1993 through December 2022, they find that: Keep Reading

Past research relating U.S. stock market returns to the party holding the Presidency mostly concludes that Democratic presidents are better for the stock market than Republican presidents. However, Presidents share power conferred by the electorate with Congress. Does historical data confirm that Democratic control of Congress is also better for stock market returns than Republican control of Congress? Is control of the smaller Senate more decisive than control of the House of Representatives? To check, we relate annual U.S. stock market (S&P 500 Index) returns to various combinations of party control of the Presidency, the Senate and the House of Representatives. Using party in power data and annual levels of the S&P 500 Index for December 1927 through December 2023 (96 years), we find that: Keep Reading

Do Federal Funds Rate (FFR) actions taken by the Federal Reserve open market operations committee reliably predict stock market and U.S. Treasuries yield reactions? To investigate, we use the S&P 500 Index (SP500) as a proxy for the stock market and the yield for the 10-Year U.S. Constant Maturity Treasury note (T-note). We look at index returns and changes in T-note yield during the one and two months after FFR actions, separately for FFR increases and FFR decreases. Using data for the three series during January 1990 through December 2023, we find that:

Keep Reading

Does publication of academic studies on stock return anomalies in the U.S. tend to extinguish these anomalies in global markets? In their November 2023 paper entitled “Does U.S. Academic Research Destroy the Predictability of Global Stock Returns?”, Guohao Tang, Yuwei Xie and Lin Zhu compare out-of-sample (post-research sample) and post-publication global returns to research-sample global returns for 87 factors described in U.S. journals. The global sample includes 38 country markets (22 developed and 16 developing). The 87 factors include those based on momentum, value, investment, profitability, intangibles and trading frictions. For each factor each month, they reform a portfolio that is long (short) the fifth of stocks expected to have the highest (lowest) next-month returns. They weight stocks in each portfolio either equally or by market capitalization according to the approach used in the associated published paper. Using data required to compute monthly returns for 87 anomalies across 38 countries with research sample end dates after 2000 during January 1990 through December 2020, they find that: Keep Reading