Can AI forecast earnings better than analysts?
A preprint tested 13 AI models on 45,964 company filings. Their earnings forecasts land near the middle of analyst forecasts and do not consistently beat individual analysts.
AI models can write an earnings forecast that looks like an analyst’s. They do not yet write one that is reliably better. In a new preprint, 13 AI models forecast next-quarter earnings per share for 45,964 company filings, and their forecasts landed, on average across models, at the 43.4th percentile of the analysts’ own forecasts [1]. They did not consistently beat individual analysts.
The numbers
- Position. The average AI forecast sat slightly below the analyst median, at the 43.4th percentile. Across the 13 models it ranged from the 36.6th to the 49.3rd.
- Accuracy. The strongest models beat the median individual analyst on roughly 40% to 43% of filings, and the best analyst on roughly 22% to 25%.
- Inside the range. Only 41.1% of AI forecasts fell inside the range of analyst forecasts for the same filing.
- Anchoring. When the models were told the analyst consensus, the inside-range share rose from 41.1% to 59.5%. The error against reported earnings went up.
What the study did
Shengkun Wang and colleagues tested 13 AI models, some general purpose and some built for finance. Each model read the management discussion from a company’s quarterly SEC filing, with extra context in some setups, and was asked for a forecast with no worked examples. The sample is 45,964 filings from 2001 to 2024. Each forecast was compared with the analyst forecasts made before the filing, and with the earnings the company later reported.
The percentile has a plain meaning. It is the share of analysts whose forecast sits at or below the model’s. A 43.4th percentile forecast is a little lower, more pessimistic, than the typical analyst’s. It says nothing about who was closer to the truth. The beat rates answer that.
This is a preprint posted on 16 October 2026. It has not been peer reviewed.
Our reading
Two results matter for anyone who buys or builds AI forecasts.
First, a forecast that looks like the consensus is not a forecast that is right. The models cluster near the analyst centre, which makes them look reasonable. Reasonable is easy to produce. Accurate is the test.
Second, anchoring to the consensus made the forecasts look safer and made them worse. More of them landed inside the analyst range, and the error against reported earnings grew. A tool that is tuned to look plausible can lose accuracy and still pass a read-through by a busy reviewer.
For a finance team, use an AI forecast as one more input to the analyst’s work, not a replacement for it. For a board or an audit committee, require a forecast log. Record each forecast with its date and method, then score it against the reported outcome each quarter, and set the threshold for trusting it before the first score comes in. For a regulator, the lesson is the same. Judge the tool on outcomes, not on how plausible its output reads.
Score forecasts against outcomes, never against plausibility.
Limits
This is a preprint on one task, models given instructions but no worked examples, and US filings. It tells us about forecasts made this way, not about every use of AI in finance.
We first posted this finding on X: https://x.com/EstimandLab/status/2115277386954756189
Sources
- Wang, S., Ji, T., Wang, L., Mansi, S., Lu, C.-T. and Chen, Y., "From SEC Filings to Financial Signals: Evaluating LLMs on Stock Market Reactions and Analyst Forecasts", arXiv 2610.24268, submitted 16 October 2026, https://arxiv.org/abs/2610.24268 (checked 2026-10-29)