The Design Principles that Informed Our New AI Forecaster
Over the past year, a number of AI forecasting tools have come to market. Broadly, they seem to be taking one of two paths:
- Querying multiple frontier models and averaging the results, or
- Training a custom forecasting model intended to maximize predictive accuracy.
Both approaches are interesting, and both optimize for something that, in real-world decision-making, often matters less than people think.
After more than a decade working as practitioners in forecasting and judgment aggregation, one lesson stands out: “superhuman accuracy” is rarely the point. What matters far more is getting an initial take early, understanding how that judgment evolves over time, seeing where there is disagreement or contrarian signal, and knowing when new information should prompt you to revisit your assumptions. As long as forecasts are good enough, they can meaningfully reduce uncertainty and materially improve decisions.
That philosophy is what shaped what we’ve just released: the Cultivate AI Forecaster.
As I wrote in our introductory post, it’s an AI-driven forecasting capability we’ve been quietly testing in live environments, including our public forecasting platforms and across hundreds of real forecast questions inside ARC. It can operate as a standalone forecasting capability or integrate directly into structured analytic workflows for scenarios, indicators, and decision support.
What makes our approach different isn’t that we ask AI models to forecast. We do. It’s how we teach them to forecast, how we structure their inputs, and what we will be doing with their outputs afterwards.
The AI Forecaster is built as a method-constrained forecasting system, not a one-off prompt or a single predictive engine. It begins with a structured research pipeline that assembles relevant evidence and context. We then apply highly-iterated instructions that are grounded in analytic tradecraft and forecasting best practices that require models to follow an explicit forecasting methodology and return structured probabilistic judgments. Those judgments are generated across multiple frontier models and then aggregated, tracked, and weighted using methods informed by our in-depth experience scoring, calibrating, and learning from human forecasts.
This design reflects a different goal. We’re not trying to win academic benchmarks or claim clairvoyance. We’re trying to support sensemaking - by being reliably good enough. In practice, the most valuable forecasts are often the ones that help teams form earlier views, avoid common biases, track how confidence shifts, and surface disagreement or contrarian thinking that sharpens analytic rigor. Often, the real utility comes from a forecast that nudges a decision-maker to pause, reconsider an assumption, or gain confidence that a direction still holds.
Forecasting failures are rarely about raw intelligence. They’re about late signals, hidden assumptions, false consensus, and overconfidence in a single perspective. Single-model systems risk hard-coding those failures and accuracy-at-all-costs to claim benchmark and forecasting competition supremacy approach risk optimizing for the wrong outcome entirely.
We built the AI Forecaster to be method-driven rather than model-centric, decision-supportive rather than performative, and complementary to human judgment rather than a replacement for it. The aim is earlier signal, clearer disagreement, and better-timed reconsideration, especially when time, attention, and staffing are constrained, and susceptibility to recency bias is high.
Want to try it out? You can sign up at https://arcanalysis.ai. And as always, I’m happy to hear feedback! Feel free to reach out directly.
