Frontier AI Models Surpass Human Experts in Earnings Forecasting for the First Time

Financial forecasting has long been regarded as the pinnacle of knowledge work. With global equity markets exceeding $150 trillion in total market capitalization, the quarterly revenue, gross margin, operating income, and adjusted EPS reported by tens of thousands of companies drive trillions of dollars in investment decisions. For decades, accurately predicting these figures depended on the years of experience of senior analysts, and whether AI could reach superhuman levels in this domain has remained an open question.

A new study from the Samaya research team provides a clear answer: frontier AI models paired with a specialized financial forecasting framework have, for the first time, surpassed human experts in predicting earnings surprises.

What Is an Earnings Surprise, and Why Is It Hard to Predict

Before each earnings release, multiple analysts covering a company each publish their own forecasts for key metrics in the upcoming quarter, and the median of these forecasts is known as the consensus estimate. When a company's actual reported figure comes in above consensus, it is called a "beat"; when it falls below, it is called a "miss." The gap between the two is the earnings surprise.

The market's reaction to earnings surprises is often extremely volatile. Precisely because predicting earnings is so difficult—even the best analyst consensus frequently deviates from actuals—earnings surprises make an ideal test case for evaluating AI's financial forecasting capabilities. Not only does it provide a clear ground truth (the actual reported figure) and a strong human benchmark (analyst consensus), but the performance drivers disclosed by management also allow researchers to examine whether the model's reasoning process is correct.

Consensus estimate is the median forecast of the sell-side analyst community. Analysts at each brokerage firm provide independent forecasts for next-quarter revenue, EPS, and other metrics based on their own financial models, industry research, and company communications. Data providers such as Refinitiv and FactSet aggregate these forecasts and take the median as the market benchmark. This mechanism means that consensus inherently carries a "correction for optimistic bias"—analysts often give forecasts slightly above their own true judgment to maintain relationships with companies, while management also guides the market to lower expectations before earnings, causing actual figures to systematically exceed consensus. Understanding this bias structure is the prerequisite for evaluating the true capabilities of any forecasting system (whether human or AI), and it is also why simply "beating consensus" is not sufficient to prove exceptional ability.

Building the Test Environment: Putting AI and Analysts on Equal Information Footing

The Samaya team built a finance-specific forecasting harness, with the core design principle of having AI models compete with human analysts under completely identical conditions.

Each forecasting task is defined as an instance: one week before an earnings release, the model must predict four core metrics for a company—revenue, gross margin, operating income, and adjusted EPS—based on all publicly available information as of that point in time.

The framework's key technical components include:

• Point-in-time gate: Enforcing information cutoffs at the framework level through authentication tokens, preventing models from accessing future data and fundamentally eliminating "cheating"
• Time-aware retrieval: A modified retrieval stack that strictly adheres to time cutoffs, with all structured data tools subject to the same restrictions
• Alternative web access: Because it is difficult to impose time gating on the open web, the team removed web access permissions and replaced them with important news sources from before the cutoff point
• Expert-guided instructions: Optimizing context efficiency and adding expert-guided reasoning instructions to improve forecast quality

The test covered 456 companies across all industry sectors, all with market capitalizations exceeding $5 billion and at least 8 brokerages providing forecasts. All of these companies released earnings after all models' knowledge cutoff dates, ensuring that models could not rely on parametric memory.

The study tested seven frontier models, including GPT-6 Astra, Claude Fable 5.1, Claude Opus 5.5, GPT-5.6 Sol, Claude Sonnet 5, Gemini 3.8 Flash, and Kimi K3.

Point-in-time gate is a core infrastructure concept in quantitative backtesting and financial AI evaluation. In real investment decisions, the information available at any point in time is strictly limited by the constraint of "what had already been made public at that time"; if a model can access future earnings data or price movements during backtesting, it produces severe look-ahead bias, making results artificially inflated and non-reproducible. By enforcing isolation of future data at the framework level through authentication tokens and embedding time filters into the retrieval stack, Samaya is essentially replicating the "as-of" semantics of professional quantitative research. The significance of this design is that it transforms "whether the model cheated" from an unauditable internal model behavior into a verifiable engineering constraint at the framework level, which is the key to the credibility of such evaluations.

Core Findings: Stronger Capabilities, More Accurate Forecasts

The research team established three evaluation metrics: forecast error measures how far predictions deviate from actuals, surprise correlation measures whether the model's predicted surprise correlates with the actual surprise, and hit rate measures whether the prediction and actual fall on the same side of consensus. All metrics are standardized by each company's historical surprise volatility.

To establish a more difficult human benchmark, the team also constructed a bias-corrected consensus—adding each company's historical median surprise to the original consensus. This is because analysts systematically depress expectations before earnings, causing actual figures to exceed consensus more often than not, so simply beating the original consensus may just be "gaming the system."

The results show a clear trend line: the stronger the model, the smaller the forecast error and the higher the surprise correlation. Earlier models such as Claude Sonnet 5 and Kimi K3 could already beat the original consensus, but only the three most recent models—Fable 5.1, Opus 5.5, and GPT-6 Astra—could beat the stronger bias-corrected benchmark.

GPT-6 Astra performed best on most metrics, including revenue forecast error, overall forecast error, and hit rate; Fable 5.1 and Opus 5.5 led on surprise correlation.

In terms of hit rate, the best AI models performed close to the historical majority benchmark on the beat direction, but achieved significant improvement on below-consensus predictions—precisely the part that the market finds hardest to anticipate.

Bias-corrected consensus is a benchmark that is harder to beat than the original consensus. Because analysts systematically lower expectations, actual reported figures exceed the original consensus more often than not, so a model that always predicts a "beat" can achieve a decent hit rate against the original consensus without truly predicting the direction of the earnings surprise. Bias correction works by adding each company's historical median earnings surprise to the original consensus, producing an adjusted benchmark where "expectations already incorporate historical bias." After this adjustment, a model must further judge the specific direction of this quarter's surprise beyond the original consensus to truly outperform, avoiding the opportunistic behavior of "free-riding" on statistical regularities. This benchmark design is methodologically critical when evaluating financial forecasting capabilities.

AI's Reasoning Process: Thinking Like a Top Analyst

The research team analyzed the reasoning traces of Claude Fable 5.1 and GPT-6 Astra, finding that the models can calculate the financial impact of world events and news like seasoned analysts. A key observation is that the models' willingness to actively gather new evidence and deviate from consensus is the core driver of their ability to outperform consensus.

In ablation studies, the importance of data access was further quantified:

• Without any data and relying solely on parametric memory, even the best models perform poorly
• When provided with data from the beginning of the quarter, performance improves by about 25 percentage points
• But there remains a 12-16 percentage point gap from using the latest data, roughly equivalent to the performance difference between an average model and a frontier model

Expert guidance is similarly effective: adding expert-guided instructions increases the model's search time and context usage by about 1.6 to 2.7 times, with forecast error decreasing accordingly.

Another interesting finding came from the "take away consensus" experiment: when all structured estimate data and text containing consensus information were removed, weaker models' performance dropped significantly, but GPT-6 Astra's performance remained almost unchanged—it may have reconstructed consensus from public information.

Ablation study is a standard experimental method in machine learning for identifying the contribution of each system component, by removing or replacing a module one at a time and observing changes in overall performance, thereby quantifying that module's necessity. Samaya's ablation design here targets two types of variables: data access (from no data at all, to only early-quarter data, to the latest real-time data) and expert-guided instructions (with/without). The value of this experiment is that it decomposes the vague question of "can AI predict" into the measurable question of "what factors affect forecast quality and to what degree," and reveals that the magnitude of data timeliness impact (12-16 percentage points) is comparable to generational model differences, suggesting that in financial forecasting, the importance of information infrastructure may be no less than the model's own parameter scale.

Difficulty Differences Across Metrics

Models performed relatively better on revenue and gross margin than on operating income and adjusted EPS. The team proposed two explanations: operating income and EPS are downstream results of the first two metrics, so errors may accumulate; at the same time, these two are adjusted figures, and different companies handle non-recurring items with varying conventions, increasing forecasting difficulty.

The study also examined the randomness of model outputs. After running 100 companies five times each, it found that although there were differences between individual runs, averaging across multiple runs did not bring significant improvement.

What This Means for the Financial Industry

The signal from this research deserves serious attention from practitioners. It is not "AI can help analysts do more," but rather "AI has already begun to work alongside analysts—and even surpass them—on core forecasting tasks."

The Samaya team stated that an alpha version of earnings forecasting is already available in the product for users and clients, and they are exploring adapting forecasts to clients' internal data to generate institution-specific insights. Research directions also include: deeply understanding the model's reasoning mechanisms, determining whether models can reliably identify true performance drivers, and using reinforcement learning post-training to enable models to continuously learn from past forecasting mistakes.

A noteworthy detail is that the forecasting harness this research relies on is itself one of the core variables. The same model may perform vastly differently under different frameworks—meaning that competition in financial AI is not only competition in model capabilities, but also competition in data infrastructure and domain knowledge engineering.

---

Note: All data in this article comes from experimental results published by the Samaya research team. Model names are as given in the original text.

One-line Summary

Frontier AI models surpass the bias-corrected human analyst consensus benchmark in earnings surprise forecasting for the first time.

Paragraph Summary

The Samaya research team built a finance-specific forecasting framework that, under the constraints of point-in-time gating, time-aware retrieval, and expert-guided instructions, had seven frontier models predict revenue, gross margin, operating income, and adjusted EPS for 456 companies one week before earnings releases. The results show a clear inverse relationship between model capability and forecast error: earlier models could already beat the original analyst consensus, while only GPT-6 Astra, Claude Fable 5.1, and Opus 5.5 could beat the stronger bias-corrected benchmark. The study further quantified the roles of data timeliness, expert guidance, and consensus information, finding that real-time data is the most important variable for accuracy, while GPT-6 Astra's performance remained almost unchanged after consensus information was removed, suggesting it can reconstruct market consensus.