Crystal ball showing a future city skyline with financial charts overlaid, symbolizing AI forecasting.

ChatGPT for Forecasting: Can AI Predict the Future Without Extra Help?

Kai Mendoza monogram
Kai Mendoza in Tech & Innovation October 2026 • 4 min read.

"Explore how basic ChatGPT models use storytelling to forecast trends and events with surprising accuracy."


Artificial intelligence is rapidly evolving, pushing the boundaries of what machines can do. Large language models (LLMs) like OpenAI's ChatGPT-4 can now mimic human speech and perform complex tasks. These predictive machines may even provide new forecasting tools, but how accurate are they?

A new study explores the forecasting abilities of ChatGPT-3.5 and ChatGPT-4, focusing on how different prompting strategies impact accuracy. By testing the models' ability to predict events from 2022, using training data only up to September 2021, the researchers uncovered surprising insights into AI's predictive potential.

The researchers used two main prompting strategies: direct prediction and future narratives. Direct prediction prompts asked ChatGPT to forecast events directly. Future narratives involved ChatGPT telling fictional stories set after September 2021, with characters sharing events that had already happened. This approach tapped into the model's creativity to see if storytelling could unlock better predictions.

AI Search Multiple angles on this topic

AI Assistants and Forecasting

ChatGPT is presented as a general-purpose AI service that can answer questions, complete work, write, create images, and code in one place. Google Cloud describes artificial intelligence as technologies that let computers learn, reason, understand language, analyze data, and provide helpful suggestions. OpenAI says its research may eventually lead to artificial general intelligence capable of solving human-level problems. Google Gemini similarly offers assistance with writing, planning, and brainstorming, illustrating how generative AI tools are positioned for a broad range of tasks.

Limits of Standalone Forecasting

Forecasting with an AI system alone should be treated cautiously because predictions depend on the information, assumptions, and context available to the system. A model may generate a plausible answer without having reliable evidence about future events. Human review, current data, and clearly defined forecasting methods may therefore remain important. These limitations mean that ChatGPT should not automatically be treated as an independent source of certainty about the future.

From AI Concepts to Generative Tools

IBM defines artificial intelligence as technology that enables computers and machines to simulate human learning, comprehension, problem solving, decision-making, creativity, and autonomy. This description places modern forecasting assistants within a broader effort to reproduce capabilities traditionally associated with human intelligence. Google AI Studio represents a more recent development-oriented environment, offering tools for building with Gemini, Veo, and Nano Banana. Together, these examples show a progression from defining AI capabilities to providing platforms for creating applications with generative models.

How Storytelling Enhances AI Forecasting

Crystal ball showing a future city skyline with financial charts overlaid, symbolizing AI forecasting.

The study found that future narrative prompts significantly improved ChatGPT-4's forecasting accuracy. When ChatGPT-4 was asked to tell stories about future events, its predictions became more accurate, especially for major Academy Award winners and economic trends. Direct prompts, on the other hand, often resulted in poorer performance.

Researchers analyzed 100 prompts and discovered that narrative prompts allowed ChatGPT-4 to leverage its ability to construct hallucinatory narratives. This capacity for narrative construction facilitated more effective data synthesis and extrapolation than straightforward predictions. It suggests that framing a request within a story helps AI process information more effectively.

  • Academy Awards: Future narratives significantly enhanced ChatGPT-4's ability to predict major Academy Award winners.
  • Economic Trends: Economic forecasts improved when the model impersonated public figures like Federal Reserve Chair Jerome Powell in narrative scenarios.
  • Falsification Exercise: Repeating experiments with more recent training data dramatically improved accuracy when the training window included the events being prompted for, achieving 100% accuracy in many instances.
AI Search Multiple angles on this topic

Evidence Before Prediction

Recent evaluations of AI forecasting should be read with attention to the data, benchmarks, and conditions used to produce the results. A system may perform well in one forecasting setting while struggling when circumstances change or information is incomplete. Reviews are therefore likely to distinguish between generating a plausible forecast and producing a consistently accurate one. Without specific evidence, no general conclusion about ChatGPT's predictive performance should be assumed.

Why Forecasts Can Fail

AI-generated forecasts can fail when the underlying information is outdated, ambiguous, or insufficient for the question being asked. Fluency may also make an uncertain prediction sound more confident than the available evidence warrants. These risks are especially relevant when users treat a model's response as a prediction rather than as an aid to reasoning. Independent checking and human judgment can help identify unsupported assumptions before action is taken.

Comparing Forecasting Tools

Comparing ChatGPT with other AI assistants requires more than comparing how naturally each system writes. Meaningful comparison would consider the quality of the source data, the clarity of uncertainty, performance on the same forecasting tasks, and how each tool responds to changing evidence. Different systems may be stronger for different uses, such as planning, explanation, or structured analysis. Without a common evaluation, claims that one assistant can predict the future better than another remain provisional.

In a falsification exercise conducted in May 2024, after the models had access to more recent training data, ChatGPT-4's accuracy dramatically improved. When the training window included the events being prompted for, the model achieved 100% accuracy in many instances. This suggests that ChatGPT-4's predictions in the initial experiments were based solely on its training data.

The Future of AI Predictions

This research highlights the potential of LLMs for predictive analysis and suggests that narrative prompts could be a key to unlocking more accurate AI forecasts. By framing predictive tasks within creative storytelling, it may be possible to access more sophisticated predictive capabilities, even in sensitive areas where direct forecasting might breach ethical considerations. This work opens new avenues for applying LLMs in economic forecasting, policy planning, and more, challenging how we interact with sophisticated AI models.

AI Search Multiple angles on this topic

AI as Forecasting Support

The available descriptions portray ChatGPT and related systems as broad-purpose tools for answering questions, creating content, and supporting work rather than as guaranteed predictors of future events. Their value in forecasting may lie in helping users organize information, explore possibilities, and articulate assumptions. Reliable conclusions still depend on evidence and the judgment applied to the output. AI is therefore best understood as a forecasting aid whose answers require evaluation.

Toward More Capable AI

OpenAI states that its research may eventually lead to artificial general intelligence capable of solving human-level problems. That ambition suggests a future in which AI systems could handle broader and more complex reasoning tasks than current assistants. However, the statement describes an eventual research goal, not a demonstrated forecasting capability. Future progress should therefore be judged by tested performance rather than by the ambition of the technology's stated direction.

AI's Wider Social Role

Google AI says it is committed to enriching knowledge, solving complex challenges, and helping people grow through useful AI tools and technologies. This frames AI development as a social and practical project, not merely a contest to produce better predictions. Systems used for forecasting may consequently affect how people understand information, make decisions, and respond to uncertainty. Their broader impact will depend on how these tools are designed, deployed, and governed.

Human Judgment Still Matters

Forecasting affects real decisions, so users must remain responsible for interpreting an AI system's output. Human experience can supply context that a prompt or dataset does not capture, while judgment can question assumptions and recognize consequences. AI may broaden the scenarios people consider, but it does not remove the need for accountability. The practical value of a forecast therefore depends on the collaboration between the tool and the person using it.

What inflation benchmarks and forecasting strategies show

In the period examined, language-model inflation forecasts had lower mean squared errors than forecasts from the Survey of Professional Forecasters. That is a comparison on this measure and in this setting, not evidence that language models forecast better in every situation.[1]

A proposed reason forecasting strategies can mislead is that models may guess events are unlikely because that answer often fits prediction datasets. Such a tendency can make benchmark performance a poor guide to the models’ underlying forecasting ability.[2]

How forecast results compare across settings

For real-time U.S. inflation forecasts, out-of-sample predictions were described as largely inaccurate and stale, while predictions in pseudo out-of-sample settings were comparable to existing benchmarks. The contrast shows that results depend on how forecasts are produced and evaluated.[3]

A separate analysis identifies temporal leakage and the difficulty of carrying evaluation results over to real-world forecasting as broad problems. These concerns make benchmark results harder to interpret as evidence of performance beyond the evaluation conditions.[4]

Forecasting-strategy evaluations have used real-world events, human predictions, and an accuracy metric. That describes a different evaluation design from inflation forecasting, so the reported results are not directly interchangeable.[2]

A forecast can be framed as though it were being made at a past date, with analysis restricted to information available by then. This setup clarifies what information a conditional forecast is allowed to use.[1]

What remains unresolved about ChatGPT forecasts

  • How well do results from a restricted, past-date forecasting setup carry over to forecasts made under real-world conditions? The evidence identifies extrapolation from evaluation performance as a concern, but does not settle that question.[1] [4]
  • How much does access to newer information change a model’s ability to produce conditional forecasts? The cited comparison notes that GPT-4’s dataset did not extend beyond September 2021 in the version discussed, while PaLM’s dataset was typically only a few days old.[1]

Sources

  1. Artificial Intelligence and Inflation Forecasts (www.stlouisfed.org)
  2. Can Language Models Use Forecasting Strategies? (research.google)
  3. ChatMacro: Evaluating Inflation Forecasts of Generative AI | Fed in Print (www.fedinprint.org)
  4. Pitfalls in Evaluating Language Model Forecasters (proceedings.iclr.cc)

About this Article -

Written with AI assistance from published research, and reviewed by the Mystum team. See our About page for more information.

Everything You Need To Know

1

How can ChatGPT be used for forecasting trends and events?

Basic ChatGPT models, specifically ChatGPT-3.5 and ChatGPT-4, can be used for forecasting by employing different prompting strategies. The models use storytelling or 'future narratives' to predict events and trends. Instead of directly asking for a prediction, ChatGPT is prompted to create a fictional story set in the future where characters discuss events that have already occurred. This method taps into the model's ability to construct narratives, leading to surprisingly accurate predictions, especially when using ChatGPT-4.

2

What are direct prediction prompts, and how do they compare to future narrative prompts when using ChatGPT for forecasting?

Direct prediction prompts involve directly asking ChatGPT to forecast an event. In contrast, future narrative prompts ask ChatGPT to create a story set after a specific date, where characters discuss events that have already happened. Research indicates that future narrative prompts significantly improve ChatGPT-4's forecasting accuracy compared to direct prompts. The narrative approach allows ChatGPT-4 to better leverage its ability to synthesize data and extrapolate trends, leading to more effective predictions in areas like economic trends and Academy Award winners. Direct prompts tend to yield poorer performance because they don't engage ChatGPT's narrative construction capabilities.

3

What role does storytelling play in enhancing the forecasting abilities of ChatGPT?

Storytelling, or the use of future narrative prompts, enhances ChatGPT-4's forecasting abilities by allowing the model to leverage its capacity to construct hallucinatory narratives. By framing a predictive task within a story, ChatGPT-4 can synthesize information and extrapolate trends more effectively than when using direct prediction prompts. This approach was particularly successful in predicting major Academy Award winners and economic trends. The act of creating a narrative appears to help the AI process information in a way that unlocks more sophisticated predictive capabilities.

4

How does the training data available to ChatGPT affect its ability to make accurate predictions?

The training data significantly impacts ChatGPT's predictive accuracy. A falsification exercise demonstrated that when the training window included the events being prompted for, ChatGPT-4 achieved 100% accuracy in many instances. This suggests that in initial experiments, ChatGPT-4's predictions were primarily based on the information it had already been trained on. Access to more recent and relevant data allows the model to make more informed and accurate forecasts, underscoring the importance of up-to-date training data for AI predictive models.

5

What are the potential future applications of using large language models like ChatGPT for predictive analysis, and what challenges or ethical considerations might arise?

Large language models (LLMs) like ChatGPT hold significant potential for predictive analysis in various fields, including economic forecasting and policy planning. The ability to frame predictive tasks within creative storytelling opens new avenues for accessing sophisticated predictive capabilities. However, challenges and ethical considerations may arise, particularly in sensitive areas where direct forecasting could breach ethical boundaries. Ensuring responsible and ethical use of these models will be crucial as they become more integrated into decision-making processes. Further research is needed to explore the full extent of their capabilities and address any potential risks.

Newsletter Subscribe

Subscribe to get the latest articles and insights directly in your inbox.