Objective research to aid investing decisions

Value Investing Strategy (Strategy Overview)

Allocations for July 2024 (Final)

Momentum Investing Strategy (Strategy Overview)

Allocations for July 2024 (Final)
1st ETF 2nd ETF 3rd ETF

Investing Expertise

Can analysts, experts and gurus really give you an investing/trading edge? Should you track the advice of as many as possible? Are there ways to tell good ones from bad ones? Recent research indicates that the average “expert” has little to offer individual investors/traders. Finding exceptional advisers is no easier than identifying outperforming stocks. Indiscriminately seeking the output of as many experts as possible is a waste of time. Learning what makes a good expert accurate is worthwhile.

Using AI to Predict Stock Prices

Can artificial intelligence (AI), in the form of a deep-learning neural network, outperform human analysts in predicting stock prices at an annual horizon? In the May 2024 revision of his paper entitled “Can AI Replace Stock Analysts? Evidence from Deep Learning Financial Statements”, Nathan Dong tests the ability of a 5-layer neural network to predict stock prices 12 months hence using recent financial statements, interest rates and stock prices. He does not impose any valuation formulas (such as discounted cash flow) or valuation ratios on the AI model. Instead, he lets the neural network, running on inexpensive computer equipment, devise its own approach. He splits his dataset dynamically into an in-sample training subset and an out-of-sample prediction subset. For example, to make a 12-month prediction in the first quarter of 2022, the training subset consists of 5,000 iterations based on the four quarterly financial statements from 2020 and interest rates/stock prices from 3, 6, 9 and 12 months ago. He measures AI model and human analyst accuracy based on mean squared error between target prices and actual future prices and the ability of variation in target prices to explain variation in actual future prices (R-squared). He also explores the factors that drive similarities/differences in target prices between human analysts and the AI model. Using quarterly financial statements, analyst 12-month target prices and actual prices for a broad sample of U.S. stocks as available, and contemporaneous yields for  10-year constant maturity U.S. Treasury notes and Moody’s seasoned Baa corporate bonds during 2010 through 2023, he finds that:

Keep Reading

Congressional Trade Tracking ETFs

Do funds based on holdings/trades of members of the U.S. Congress and their families beat the market? To investigate, we look at performances of two recently introduced exchange-traded funds (ETF):

  1. Unusual Whales Subversive Democratic Trading ETF (NANC) – invests primarily in stocks held by Democratic members of Congress and/or their families per public disclosure filings.
  2. Unusual Whales Subversive Republican Trading ETF (KRUZ) – invests primarily in stocks held by Republican members of Congress and/or their families per public disclosure filings.

We use SPDR S&P 500 ETF Trust (SPY) as a benchmark. Using daily dividend-adjusted prices for NANC, KRUZ and SPY during February 7, 2023 (NANC and KRUZ inception) through June 7, 2024, we find that: Keep Reading

AIs for Financial Statement Analysis?

Are large language models such as GPT-4 as effective as professional human analysts in interpreting numerical financial statements? In their May 2024 paper entitled “Financial Statement Analysis with Large Language Models”, Alex Kim, Maximilian Muhn and Valeri Nikolaev investigate whether GPT-4 can analyze standardized, anonymized financial statements to forecast direction and magnitude (large, moderate or small) of changes in future firm earnings and provide the level of confidence in its answer. They withhold management discussions that accompany financial statements, choosing to evaluate the ability of GPT-4 to analyze only numerical data. They anonymize statements by omitting firm names and replacing years with labels (t, t − 1, …) so that GPT-4 cannot use its training data to find actual future earnings. They consider both a simple query and a series of prompts designed to make GPT-4 think like an ideal human analyst by focusing on changes in certain financial statement items, computing financial ratios and generating economic interpretations of these ratios. They compare GPT-4 forecasts to: (1) consensus (median) human earnings forecasts issued during the month after financial statement release; and, (2) forecasts from other benchmarks, including that of a highly focused state-of-the-art artificial neural net (ANN) model. To test economic value of forecasts, they each year on June 30 form portfolios using GPT-4 forecasts based on annual financial statements from the preceding calendar year end, as follows: 

  • Sort stocks based on GPT earnings forecasts.
  • Select stocks expected to have moderate/large increases or decreases in earnings and separately resort these two groups based on forecast confidence.
  • Form an equal-weighted or value-weighted long (short) portfolio of the tenth, or decile, of these stocks with highest confidence in earnings increases (decreases).

Using financial statements for 15,401 firms during  1968 through 2023 (with 2022 and 2023 out-of-sample with respect to the GPT-4 training period), annual returns of associated stocks and consensus human analyst earnings forecasts for 3,152 firms during 1983 through 2021, they find that: Keep Reading

Making AI Do Numbers to Predict Stock Returns

Can large language models like ChatGPT work with numbers to support technical analysis of stock returns rather than just words to support sentiment analysis? In his April 2024 paper entitled “StockGPT: A GenAI Model for Stock Prediction and Trading”, Dat Mai introduces StockGPT, an autoregressive model trained and tested on stock returns rather than firm news. He segments 1926 through 2000 daily U.S. stock returns into intervals (tokens) and then trains StockGPT to identify predictive return patterns. He then tests the model, using daily returns over the past 256 trading days to predict next-day returns during 2001 through 2023. He assesses accuracy of return predictions by:

  1. Running cross sectional regressions of actual versus predicted daily returns.
  2. Each day at the market close: (1) removing stocks in the lowest tenth of market values; and, (2) reforming an equal-weighted or value-weighted hedge portfolio that is long (short) stocks within the highest (lowest) tenth, or decile, of predicted returns.

He performs sensitivity tests that account for portfolio costs, portfolio construction delay and elimination of low-priced stocks. Using daily returns for all common stocks traded on NYSE, AMEX or NASDAQ during 1926 through 2023, he finds that: Keep Reading

Failure of Non-causal Factor Strategies

Do widely used associational (rather than causal) methods used by researchers to specify factor models of asset returns work? In their March 2024 paper entitled “The Case for Causal Factor Investing”, Marcos Lopez de Prado, Alex Lipton and Vincent Zoonekynd describe the shortcomings of associational methods of factor model development. They address p-hacking (data snooping), with focus on interferences from variables called colliders (causally influenced by two or more variables) and confounders (influencing both dependent and independent variables). They further describe what can be done to correct these shortcomings. Based on logical/mathematical analysis and the body of financial markets research, they conclude that:

Keep Reading

ChatGPT-generated Financial News Sentiment and NASDAQ Returns

Can ChatGPT extract market sentiment from financial news that is useful for timing equity markets? In their April 2024 paper entitled “Sentiment Analysis of Bloomberg Markets Wrap Using ChatGPT: Application to the NASDAQ”, Baptiste Lefort, Eric Benhamou, Jean-Jacques Ohana, David Saltiel, Beatrice Guez and Thomas Jacquot use ChatGPT to assess whether daily Bloomberg Global Markets Wrap, Market Talks and Morning Reports anticipate NASDAQ returns. Specifically, they each day:

  • Ask ChatGPT to identify important news themes and characterize them as headlines.
  • Ask ChatGPT to assess whether each headline is positive, negative or neutral for future stock prices.
  • Compute a sentiment score that combines sentiments for all daily headlines.
  • Compute a daily cumulative sentiment score (C) for the last 20 trading days.
  • Compute a daily detrended cumulative sentiment score (DC) by comparing C to its value over the last 20 trading days (extending the overall lookback interval to 40 days).
  • If C is positive (negative), take a long (short) position in the NASDAQ index with a 2-day lag to ensure executability and a debit of 0.2% trading frictions for position changes. Repeat this evolution for DC.

They separately examine cumulative performances of the long, short and overall returns of the C and DC variations of this strategy, focusing on Sharpe, Sortino and Calmar ratios as key performance metrics. Their benchmark is buying and holding the NASDAQ index. Using the specified daily financial news sources and daily NASDAQ index returns during 2010 through 2023, they find that:

Keep Reading

Coordinated Retail Traders Won the War with Short Sellers?

Do short-selling hedge funds consistently extract alpha from exuberant retail traders? In their March 2024 paper entitled “Short-Selling Hedge Funds”, Jialin Qian, Zhen Shi and Baozhong Yang examine the performance of hedge funds engaged in short-selling, as follows:

  1. Which hedge funds are likely short-sellers, and how do they compare with other hedge funds?
  2. What factors contribute to the performance of short-selling hedge funds?
  3. How has the 2021 Meme stock phenomenon affected short-selling hedge funds?

They each month identify short-selling hedge funds as those with positive return betas over the past 24 months versus a monthly rebalanced portfolio of short stock positions with weights proportional to their respective short interests. They relate behaviors of short-selling funds to those of other hedge funds and to those of retail traders. Using monthly data for 11,054 U.S. hedge funds, returns and short interests for a broad sample of U.S. stocks and data to measure retail stock trading/sentiment during 2010 through 2022, they find that: Keep Reading

Lookahead Bias in Large Language Model Training Data

Can Large Language Models (LLM) inject lookahead bias into backtests when rigor is lacking in generation of LLM training samples? In their preliminary and incomplete March 2024 paper entitled “Lookahead Bias in Pretrained Language Models”, Suproteem Sarkar and Keyon Vafa examine the potential for lookahead bias in backtests using the Llama-2 LLM to identify future firm risks based on content of earnings calls. They consider cases for which: (1) the backtest falls within the LLM training sample, but the researcher tells the LLM to consider only information before the test period; and, (2) the researcher specifies a training sample that ends before the backtest but generates it long after the end of the training sample. Using Llama-2 to interpret transcripts of selected firm earnings calls from 2018, they find that:

Keep Reading

Informativeness of Seeking Alpha Articles for Stock Returns

Are sentiments conveyed in Seeking Alpha articles useful for stock picking? In their January 2023 paper entitled “Seeking Alpha: More Sophisticated Than Meets the Eye”, Duo Selina Pei, Abhinav Anand and Xing Huan apply two-pass natural language processing to test the informativeness of articles from Seeking Alpha incremental to publicly available earnings data. Specifically, they each month:

  • Associate articles with one or more specific stocks.
  • Extract positive and negative sentiment at both phrase and aggregate levels for each article/stock.
  • Calculate a standardized net sentiment for each article/stock based on the difference between positive and negative mentions, emphasizing event sentiment over general sentiment.
  • Rank articles/stocks based on standardized net sentiment over the last month. Reform equal-weighted portfolios of articles/stocks by ranked tenths (deciles). Calculate both immediate [-1,+1] and 90-day future [+2,+90] average gross raw returns and average gross abnormal returns adjusted for size, book-to-market and momentum.
  • Sort stocks into 20 groups based on monthly standardized net sentiments up to two days before portfolio selection, excluding stocks with few articles or neutral sentiment. Reform an equal-weighted hedge portfolio that is long stocks with the highest sentiments and short stocks with the lowest (on average, 105 long and 86 short positions).

Using 350,095 articles published on Seeking Alpha since its inception in 2004 through the beginning of October 2018, daily returns of matched stocks and their options and associated earnings surprise data as available, they find that: Keep Reading

Day Trading Stocks with ChatGPT

Can artificial intelligence platforms such as ChatGPT be good stock day traders? In his March 2024 paper entitled “Can ChatGPT Generate Stock Tickers to Buy and Sell for Day Trading?”, Sangheum Cho tests whether ChatGPT 3.5 turbo supports profitable day trading. He instructs ChatGPT to pretend to be a professional day trader who picks from among U.S. listed stocks 100 to buy and 100 to sell for short-term returns based on daily Bloomberg and the Wall Street Journal news blurbs on Twitter. Each day, prior to the market open, he:

  • Uses the Refinitiv Eikon News Monitor to collect the selected tweets from the past 24 hours. He removes hyperlinks and duplicate tweets.
  • Segments the tweets into batches to accommodate ChatGPT processing limitations.
  • For each batch, asks ChatGPT to generate 100 BUY and 100 SELL signals, with 30 iterations for each batch to amplify signals by suppressing spurious selections. He then constructs equal-weighted long and short portfolios of stocks with signals.
  • For each stock with signals:
    • Sums BUY and SELL signals across batches/iterations to calculate SUM_BUY and SUM_SELL signals. He constructs signal count-weighted long and short portfolios from these summed signals.
    • Subtracts SUM_SELL from  SUM_BUY to calculate NET_BUY and NET_SELL signals. He constructs signal magnitude-weighted long and short portfolios from these netted signals.

For each portfolio, he excludes stocks with zero daily volume, missing daily prices or incomplete trading histories for the previous five trading days. He measures returns from the market open to the market close. Using 222,659 tweets (only 16,359 of which are firm-specific) and daily opening and closing prices for U.S. listed common stocks during December 2022 through December 2023 (271 trading days), he finds that:

Keep Reading

Daily Email Updates
Filter Research
  • Research Categories (select one or more)