Investing Expertise

How well do the short-term stock picks of publicly available artificial intelligence (AI) platforms perform? To investigate, we asked Grok, ChatGPT, Perplexity, Gemini and Meta AI the following questions on April 20, 2025:

• Please succinctly provide your unique best long idea for the next 30 days.
• Please succinctly provide your unique best shorting idea for the next 30 days.

We then: (1) calculated total returns for the resulting stock picks from the close on April 21, 2025 to the close on May 21, 2025; (2) averaged the returns for long and short picks; and, (3) compared  average returns for long and short picks. We include total returns for SPDR S&P 500 ETF (SPY) and Invesco QQQ Trust (QQQ) over this same interval for context. Using dividend-adjusted prices for the specified picks, we find that:

Keep Reading

Should investors, particularly those employing machine learning, prefer complex or simple prediction models? In the May 2025 revision of his paper entitled “Simplified: A Closer Look at the Virtue of Complexity in Return Prediction”, Daniel Buncic challenges prior research finding that portfolio performance (Sharpe ratio) increases with machine learning model complexity when the number of inputs (potential predictors) greatly exceeds the number of training observations. Using the same dataset, prediction models and portfolio evaluation methods as the prior research, he finds that: Keep Reading

We occasionally ask publicly available artificial intelligence (AI) platforms for investing ideas and post results on the CXOAdvisory X account. Two recent examples are:

1. “Please concisely provide your unique choice for the best risk-adjusted investment for generating monthly or quarterly income.”
2. “Using data up to now, please concisely provide your unique estimate of which asset class will be strongest and which will be weakest over the next three months.”

We may use responses of such items to assess usefulness of AIs as investing aids, such as in a CXOAdvisory.com post scheduled for later this week.

Can large language models (LLM) be trusted for economic/financial forecasts during periods within their training data? In their April 2025 paper entitled “The Memorization Problem: Can We Trust LLMs’ Economic Forecasts?”, Alejandro Lopez-Lira, Yuehua Tang and Mingyin Zhu evaluate use of  ChatGPT 4o (knowledge cutoff October 2023) for economic/financial forecasting via:

• Forecasts of variables before and after knowledge cutoff.
• Explicit instructions to ignore knowledge during periods before the cutoff.
• Masking of inputs (anonymized firm names or dates) to mitigate use of memorized data in forecasts before knowledge cutoff.

Using data for major economic indicators, stock index levels, individual stock returns/conference calls and Wall Steet Journal (WSJ) headlines during December 1989 through February 2025, they find that: Keep Reading

Should stock return model complexity guide breadth of input data? In their May 2025 paper entitled “Model Complexity and the Performance of Global Versus Regional Models”, Minghui Chen, Matthias Hanauer and Tobias Kalsbach assess the predictive performance of global versus regional inputs for stock return models based on linear and machine learnings algorithms: ordinary least squares regression (OLS); elastic net (ENET); random forest (RF); gradient-boosted regression trees (GBRT); and, neural networks (NN). Monthly model inputs include 36 firm-level characteristics and associated stock trading data in U.S. dollars for 24 developed market countries, suppressing effects of megacaps and excluding microcaps (the smallest stocks per country comprising 3% of overall market capitalization).  They segment country markets into four regions: North America, Europe, Japan and Asia Pacific. Model training employs an expanding window (initially six years, extended year by year), followed by a 6-year validation interval and a 1-year test interval. For each model, each month, they reform a portfolio that is long (short) the fifth, or quintile, of stocks with the highest (lowest) predicted returns. Using the specified monthly firm/stock inputs during July 1990 through December 2021, they find that:

Keep Reading

Can investors exploit analyst stock price targets by finding the best analysts and overweighting the most extreme target-implied returns? In their March 2025 paper entitled “Alpha in Analysts”, Álvaro Cartea and Qi Jin test the informativeness and exploitability of sell-side analyst stock price targets. To test informativeness of target prices, they each month for each analyst:

• Use price targets to deduce 12-month return forecasts.
• Form a hedge portfolio that is long (short) stocks with positive (negative) return forecasts, with weights proportional to magnitudes of forecasted returns and absolute value of the sum of weights equal to one.
• Compare analyst portfolio performance to that of an equal-weighted, long-only portfolio of the same stocks.

To test exploitability of results, they each month:

• Predict portfolio profitability for each analyst via an inception-to-date regression of six analyst performance metrics up to 12 months ago (capturing historical performance and breadth of stock coverage) versus next-month portfolio return.
• Construct a portfolio of analyst portfolios with higher (lower) allocations to those with higher (lower) predicted returns.

Using daily analyst price targets and associated stock returns/firm characteristics as available for common NYSE/AMEX/NASDAQ stocks during January 1999 through November 2024, they find that: Keep Reading

Are more factors better for predicting stock market returns? In their April 2025 paper entitled “The Limited Virtue of Complexity in a Noisy World”, Álvaro Cartea, Qi Jin and Yuantao Shi analyze the interactions between stock factor noisiness (errors/uncertainties) and factor model complexity in portfolio optimization. Specifically, they study how factor noise affects the predictive power of factor-based return models. They postulate that noise arises from issues such as inconsistent data collection, processing errors or insufficient computing power. Their conclusions derive from three examples:

1. Employ market data to simulate ideal noiseless stock factors, and combine them via ridge regression to predict stock returns.
2. Apply a neural network to actual NYSE/AMEX/NASDAQ data to predict monthly stock returns via a varying number of factor inputs during 1991 through 2023.
3. Use simulated stock factors which have either fixed noise levels or noise levels that increase with model complexity to explore how predictive power varies with noise and complexity. 

Using monthly excess returns for the broad value-weighted U.S. stock market and values for 15 factors used in other studies during 1926 through 2023, they find that: Keep Reading

Should those aiming to exploit machine learning for portfolio construction focus model training on the broad market or specific industries? In their April 2025 paper entitled “Do Machine Learning Models Need to Be Sector Experts?”, Matthias Hanauer, Amar Soebhag, Marc Stam and Tobias Hoogteijling examine return predictability using several machine learning (ML) models trained on a comprehensive set of firm/stock characteristics in three ways:

1. Generalist – trained on all stocks in the sample.
2. Specialist – trained on stocks only within one of 12 industry classifications.
3. Hybrid – integrates overall sample and industry information via industry-neutral mappings from stock characteristics to expected returns.

They employ four ML models, including elastic nets, gradient boosted regression trees, 3-layer neural networks and an equal-weighted ensemble of the three. They train and tune these models with an expanding window with an initial 18-year training set, 12-year validation set and 1-year test set, shifted forward each year but retaining the initial training start point. Input data consists of monthly stock returns and monthly values of 153 firm-level characteristics for U.S. stocks each month at or above the 20th percentile of NYSE market capitalizations . They assign stocks to the 12 industries (including Other), with average weights ranging from 22.5% for Tech to 1.4% for Durables. They then each month sort stocks into tenths (deciles) by machine learnings ensemble-predicted next-month return and reform a volatility-scaled, value-weighted hedge portfolio that is long the decile with the highest expected returns and short the decile with the lowest. Using the specified inputs during January 1957 (January 1986 for a non-U.S. sample) through December 2023, they find that:

Keep Reading

Does the evolution of empirical asset pricing point inevitably to machine learning methods? In his February 2025 paper entitled “From Econometrics to Machine Learning: Transforming Empirical Asset Pricing”, Chuan Shi summarizes the transition from traditional methods to machine learning in empirical asset pricing. He traces the historical development of traditional asset pricing models and their roles as benchmarks for decades of research. He compares the strengths and weaknesses of traditional methods and machine learning, explaining why the latter is well-suited to address challenges of the big data era. Finally, he introduces an approach based on the stochastic discount factor (SDF), melding the simplicity of traditional models and the flexibility/predictive power of machine learning. Based on the body of research on asset pricing, he concludes that: Keep Reading

Is the ability of artificial intelligence (AI) platforms such as ChatGPT to summarize and interpret large volumes of financial data altering investor trading behaviors and thereby changing financial markets? In the April 2025 revision of their paper entitled “ChatGPT and the Stock Market”, Jenny Stanco and Kee Chung examine the impact of ChatGPT on stock trading, volatility, liquidity, and price efficiency. For their analysis, they separate firms into those with abundant publicly available information (high-info) and those with limited information (low-info), employing firm size and age as proxies for information availability. They further use Google search volumes to estimate levels of attention firms may get from ChatGPT. They use the year before (after) ChatGPT launch on November 30, 2022 as the pre-launch (post-launch) subperiod. Using daily trading volumes and return volatilities, and earnings forecasts/announcements/actuals data, for a broad sample of U.S. stocks from the end of November 2021 through the end of November 2023, they find that: Keep Reading