Table of Contents
- Summary: the seven findings that matter
- 1. Worse than predicting zero
- 2. When the random walk wins
- 3. Against a machine learning benchmark
- 4. What returns the most optimistic deliver
- 5. What if the bias were rational?
- How to apply this practically
- Frequently asked questions
- Do analysts get their earnings forecasts right?
- Is it bias or difficulty?
- When is a simple model better?
- Is there an optimistic bias, and how much does it cost?
- Does that mean forecasts are useless?
- What we do not know
- References
Analysts’ long-run earnings growth forecasts are less accurate than a naive model that simply predicts zero growth. That is not a provocation: it is the literal conclusion of an evaluation published in an accounting and finance journal. And there is more: the random walk beats analysts over long horizons and in small firms, and a naive extrapolation of their own one-year forecast is more accurate than their explicit two- and three-year forecasts. At the same time, the comparison with machine learning models concludes that analysts remain superior to the linear models actually in use. This article sets out what has been measured.
Summary: the seven findings that matter
Each row names the study and the year. It is worth separating accuracy from bias: they are two different problems.
| Finding | Result | Study and year |
|---|---|---|
| Long-run accuracy | Worse than predicting zero growth | Harris, J. Business Finance & Accounting (1999) |
| Main source of that inaccuracy | Random error, not bias | Harris (1999) |
| Where the random walk wins | Long horizons, small or young firms | Bradshaw et al., Review of Accounting Studies (2012) |
| The most accurate two- and three-year forecast | Naive extrapolation of the one-year one | Bradshaw et al. (2012) |
| Direction and evolution of the bias | Optimistic, and growing with the horizon | Van Binsbergen et al., Rev. Financial Studies (2022) |
| Comparison with linear statistical models | Analysts are superior to them | Van Binsbergen et al. (2022) |
| Returns of the most optimistic stocks | Lower than those of the most pessimistic | Bordalo et al., The Journal of Finance (2019) |
What you will find here
• The comparison with the dumbest possible model, and who wins.
• The precise conditions under which the random walk beats an analyst.
• What happens when analysts are compared with machine learning.
- What returns the stocks with the most optimism around them deliver.
• The argument that the optimistic bias may be rational rather than a mistake.
1. Worse than predicting zero
Richard Harris analyzed how accurate, biased, and efficient long-term earnings growth predictions made by analysts for American companies really are, publishing his findings in Journal of Business Finance & Accounting. The opening sentence of his results gets straight to the point: forecast accuracy is shown to be extremely low.
And the comparison that makes it memorable: analysts’ long-run earnings growth forecasts are less accurate than those of a naive model in which earnings growth is forecast to be zero.
The paper adds that, consistently with short-term and intermediate forecasts, long-run forecasts are both biased and inefficient, and that there is evidence that analysts do not fully incorporate the information about future earnings contained in the current share price.
The nuance almost nobody quotes
Here is the part that prevents the finding from becoming a caricature: the bias and inefficiency of the forecasts contribute very little to their inaccuracy, which turns out to be mainly a consequence of random error.
Put another way: the problem with analysts over the long run is not that they push the figures in one direction; it is that a company’s earnings five years out are largely unpredictable. The study itself identifies where they do work: the most reliable growth forecasts are low forecasts issued for large firms with a low price-earnings ratio and a high market-to-book ratio.
| Study | What it analysed | What it found | What it does not support |
|---|---|---|---|
| Harris (1999) | Long-run growth forecasts in the US | Less accurate than predicting zero growth | It is long-run growth, not annual forecasts |
| Bradshaw et al. (2012) | Annual forecasts against the random walk | The random walk wins in specific horizons and firms | Does not assess the usefulness of forecasts for investing |
| Van Binsbergen et al. (2022) | Analysts against a machine learning benchmark | Growing optimistic bias, but better than linear models | The statistical benchmark is not observable in real time without a model |
| Bordalo et al. (2019) | Portfolios sorted by forecast optimism | The most optimistic return less than the most pessimistic | It is a test of a belief-formation model |
| Lim (2001) | Analyst utility model and individual data | The optimistic bias can be optimal and rational | It is a model with empirical testing, not an experiment |
2. When the random walk wins
Mark Bradshaw, Michael Drake and two co-authors re-examined in Review of Accounting Studiesthe widely held belief that analysts’ earnings per share forecasts are superior to random walk time-series forecasts.
Their result delimits precisely where that belief fails: simple random walk forecasts are more accurate than analysts’ forecasts over longer horizons, in smaller or younger firms, and when analysts forecast negative or large changes in earnings per share.
And they add a comparison that is almost comic: they also contrasted an extra one-third-longer earnings projection grounded in a simplistic extrapolation of analysts’ one-year forecasts. Surprisingly, that naive extrapolation delivers the most precise estimation of long-term earnings spanning two and three years.
The authors’ conclusion serves as a methodological caution: these insights transform previous generalisations regarding the supremacy of analysts’ predictions, indicating they are incomplete, deceptive, or both.
3. Against a machine learning benchmark
Jules van Binsbergen, Xiao Han and Alejandro Lopez-Lira introduced in The Review of Financial Studies a different way of measuring the problem: a real-time measure of the conditional biases in earnings forecasts, defined as the difference between analysts’ expectations and a statistically optimal, unbiased benchmark built with machine learning.
By that metric, the assessment regarding prejudice is unmistakable: conditional analyst expectations tend on average to be overly optimistic, a distortion that grows as the prediction timeline lengthens.
| Observation | What it shows | Practical use |
|---|---|---|
| Direction of the bias | Upward, on average | Discount optimism by default |
| Evolution with the horizon | Increases the further out you look | Three-year forecasts are the worst |
| Relation to future returns | Negative cross-sectional predictability | More optimism, lower returns afterwards |
| Where the most biased cluster | In the short legs of many anomalies | Matches what funds sell short |
| What their managers do | More likely to issue equity | They exploit the high valuation |
Yet the very same study concludes with an important defense of analysts that deserves mention: standard linear earnings models fail out of sample, underperforming the forecasts generated by financial analysts. Put differently, analysts fall short of an ideal machine learning benchmark while still outperforming current statistical alternatives.
4. What returns the most optimistic deliver
Pedro Bordalo, Nicola Gennaioli and two co-authors revisited a classic finding in The Journal of Finance and explained it with a belief-formation model. The starting fact: the returns of stocks with the most optimistic analyst long-run earnings growth forecasts are lower than those of stocks with the most pessimistic forecasts.
Their explanation rests on the representativeness heuristic: analysts forecast fundamentals from observed earnings growth, but overreact to news by exaggerating the probability of the states that have become more likely.
The authors document the joint dynamics of fundamentals, expectations and returns of those portfolios, and note that a quantitative estimation of the model accounts for the main patterns in the data.
5. What if the bias were rational?
Terence Lim proposed in Journal of Finance the reading that prevents optimism from being treated as simple incompetence. His model posits a quadratic loss utility function in which analysts trade off bias for better access to company management and for accuracy.
Something rather counter-intuitive emerges here: the best possible predictions featuring the lowest expected error tend to be optimistically skewed, displaying predictable cross-sectional differences tied to both firm and analyst traits.
And the empirical testing with individual analysts’ forecasts is consistent with the model’s predictions, hence the conclusion: a positive and predictable bias may be a rational property of optimal earnings forecasts, and prior studies using classical notions of unbiasedness may have prematurely dismissed analysts’ forecasts as irrational or inaccurate.
That logic fits an independent observation: there is more demand for non-public information about firms whose earnings are hard to predict, and if optimism eases access to management’s non-public information, analysts will issue more optimistic forecasts precisely for those firms.
How to apply this practically
Ordered by the strength of the evidence behind each point:
• Ignore five-year growth forecasts. They are less accurate than assuming zero growth.
• Discount optimism, and more the further out you look. The upward bias increases with the forecast horizon.
- Be skeptical of general agreement regarding tiny or emerging companies. This is precisely where unpredictable market behavior outperforms expert forecasts.
- Be wary of the portfolio the consensus loves most. Stocks with the most optimistic forecasts have returned less than the most pessimistic ones.
• Do not throw away the one-year forecasts. They remain better than the linear models in use, and extrapolating them beats the explicit three-year forecast.
Frequently asked questions
Do analysts get their earnings forecasts right?
Over the long run, very little. The evaluation of the accuracy, bias and efficiency of long-run earnings growth forecasts for US firms concludes that accuracy is extremely low and that those forecasts are less accurate than a naive model in which growth is forecast to be zero.
Is it bias or difficulty?
Mainly difficulty. The same study documents that the forecasts are biased and inefficient, but that bias and inefficiency contribute very little to their inaccuracy, which is mainly a consequence of random error. The most reliable are low forecasts for large firms with a low price-earnings ratio and a high market-to-book ratio.
When is a simple model better?
In three specific, measured situations, simple random walk forecasts outperform analyst predictions over extended horizons, for smaller or younger companies, and whenever analysts anticipate negative or substantial shifts in earnings per share. Furthermore, a straightforward extrapolation of the one-year projection delivers the most precise estimate of earnings at the two- and three-year marks.
Is there an optimistic bias, and how much does it cost?
There is, it grows with the horizon and it has market consequences. Measured against an unbiased benchmark built with machine learning, analysts’ conditional expectations are on average upward biased, a bias that increases with the horizon, is associated with negative cross-sectional return predictability, and clusters in the short legs of many anomalies.
Does that mean forecasts are useless?
No, and there are two reasons to qualify it. The same paper that documents the bias concludes that commonly used linear earnings models do not work out of sample and are inferior to those provided by analysts. And the model published in Journal of Finance argues that a positive, predictable bias may be a rational property of optimal forecasts, so classical studies may have dismissed them prematurely.
What we do not know
• How much of the bias is incentive and how much is belief. Rational models and overreaction models explain the same data.
- Whether machine learning improves real decisions. The unbiased benchmark is built after the fact with a model; using it in real time is another problem.
• How this carries over outside the United States. Much of the evidence comes from US markets with dense analyst coverage.
• To what extent is the discount already factored into the consensus. If market participants realize that projections are overly optimistic, that bias ought to be partially reflected in the valuation.
References
1. Harris, R. D. F. The Accuracy, Bias and Efficiency of Analysts’ Long Run Earnings Growth Forecasts. Journal of Business Finance & Accounting, 1999. DOI: 10.1111/1468-5957.00273
2. Bradshaw, M. T., Drake, M. S., Myers, J. N. and Myers, L. A. A re-examination of analysts’ superiority over time-series forecasts of annual earnings. Review of Accounting Studies, 2012. DOI: 10.1007/s11142-012-9185-8
3. Van Binsbergen, J. H., Han, X. and Lopez-Lira, A. Man versus Machine Learning: The Term Structure of Earnings Expectations and Conditional Biases. The Review of Financial Studies, 2022. DOI: 10.1093/rfs/hhac085
4. Bordalo, P., Gennaioli, N., La Porta, R. and Shleifer, A. Diagnostic Expectations and Stock Returns. The Journal of Finance, 2019. DOI: 10.1111/jofi.12833
5. Lim, T. Rationality and Analysts’ Forecast Bias. Journal of Finance, 2001. DOI: 10.1111/0022-1082.00329
This material is purely informational and does not amount to financial guidance or an investment proposal. The referenced research examines distinct markets and timeframes which cannot be equated, and none of these studies permits the dependability of any specific prediction to be foreseen.