Machine vs. market: quantifying investor belief distortions
Francesco Bianchi, Do Q Lee, Sydney C. Ludvigson, and Sai Ma introduce in this paper the concept of “prestakes”—predictable mistakes in financial markets—and use machine learning to quantify systematic errors in investor beliefs. The authors demonstrate that stock market participants make systematic, predictable mistakes that represent meaningful departures from rational and efficient expectation formation. These prestakes are especially prevalent during market turbulence when stakes are highest, suggesting real-world market beliefs contain substantial distortions that could be mitigated with better information processing.
Key methodological components:
Machine learning algorithm
The authors use Long Short-Term Memory (LSTM) neural networks trained on data from 1970-2005, then tested on 2005-2023. The algorithm independently learns from thousands of real-time economic time series including macro fundamentals, financial market data, credit risk variables, Fed funds futures, and media sentiment from Wall Street Journal articles. Critically, the machine is structured to deviate from survey forecasts only when it detects predictable mistakes in hold-out validation samples, using exclusively information that would have been available to investors at forecast time
Objective vs. subjective beliefs
Objective beliefs are defined as machine forecasts that remove human behavioral biases and efficiently process large information sets. Subjective beliefs, on the other hand, are derived from survey data from investors (UBS/Gallup, Michigan, Conference Board), CFOs, professional forecasters (Livingston), and equity analysts (IBES). Period-specific bias (prestakes) is measured as: Bias = Survey Forecast – Machine Forecast.
Findings
Substantial forecast accuracy gaps
- Machine forecasts are 25-35% more accurate than investor surveys for one-year-ahead stock returns
- Machine is 64% more accurate than IBES analysts for one-year-ahead earnings growth (70% during the GFC period)
- Survey respondents failed to predict negative returns during the 2008 financial crisis and 2022 bear market, while the machine successfully forecasted these downturns
Sources of predictable mistakes
Three key differences between subjective and objective beliefs:
- Insufficient adaptation: Subjective beliefs vary too little and don’t adapt quickly or non-linearly enough to changing conditions
- No learning from mistakes: Little evidence that survey respondents learn from past errors
- Recency bias: Forecasts largely explained by local means, even during extreme market turbulence
- Meanwhile, the machine accesses long-memory functionality, paying greater attention to credit risk, Fed funds surprises, consumer spending, and media sentiment to identify turning points during turbulent times.
Risk-return tradeoff distortions
Objective risk premia move 2 to 6.8 times more with objective risk measures than subjective risk premia move with subjective risk measures—suggesting investors are sub-optimally responsive to changing risk.
Trading strategy performance
Strategies betting on objective beliefs (implicitly betting against prestakes) earn large risk-adjusted returns. By contrast, strategies based on subjective beliefs earn zero or negative alphas because they virtually never anticipate losses.
Correlation evidence
Four of six investor surveys are negatively correlated with subsequent realized returns. Machine forecasts have 0.56 correlation with realized returns while Survey bias (prestakes) shows -0.37 to -0.70 correlation with subsequent returns, with subjective beliefs most inaccurate when excessively optimistic.
The Prestakes of Stock Market Investing
Authors: Francesco Bianchi, Do Q Lee, Sydney C. Ludvigson, Sai Ma
From: Johns Hopkins University, New York University, Federal Reserve Board
Learning from machines: belief distortions and employment dynamics
In this paper, Do Lee focuses on how belief distortions in corporate earnings expectations translate into real labor market outcomes, unlike Bianchi et al.’s paper above which concentrates exclusively on stock market returns and the efficiency of return forecasting. Lee’s paper examines both survey forecasts from equity analysts and CFO expectations, connecting expectational errors to firm hiring behavior and unemployment dynamics. Similarly to Biancho et al., Lee also employs a machine learning benchmark approach to measure belief distortions in corporate hiring and asset pricing decisions.
The author compares survey forecasts (from equity research analysts and CFOs) against machine learning forecasts (using Long Short-Term Memory neural networks trained on real-time macroeconomic and financial data). The analysis combines quarterly data from 2005–2023 covering S&P 500 firms with IBES analyst coverage, labor market metrics (vacancies, unemployment, hiring rates), and survey expectations on earnings growth and stock returns.
Findings
- Primary evidence of belief distortions: Survey forecasts systematically overreact to cash flow news, while machine forecasts do not. Positive forecast revisions are followed by negative forecast errors, indicating excessive optimism in subjective beliefs.
- Quantitative impact on labor markets: Belief distortions explain over 60% of variation in hiring at both aggregate and firm levels. They account for 90% of vacancy filling rate variation and 68% of unemployment variation in time-series analysis and they explain 66% of dispersion in firm-level hiring rates in cross-sectional analysis.
- Differential firm behavior: Following positive idiosyncratic cash flow shocks, firms with distorted beliefs overhire relative to firms with objective beliefs, leading to lower subsequent profits per worker.
- Model results: A calibrated search-and-matching model with fading-memory learning replicates empirical patterns and explains 60% of U.S. unemployment volatility and 66% of hiring dispersion—substantially more than standard rational models which explain only 16-30%.
- Asset Pricing Implications: The same belief distortions that drive labor market fluctuations also predict stock returns, connecting real economic activity to financial market dynamics through a unified framework of distorted expectations about future cash flows
Belief Distortions, Asset Prices, and Unemployment Fluctuations
Author: Do Q Lee
From: New York University
