CastFSR: A Fast--Slow--Reflect Agentic Reasoning Framework for Context-Aware Time Series Forecasting

Imagine you are trying to guess what the weather will be like tomorrow. You could just look at the calendar and say, "It's July, so it will probably be hot," or you could look at the thermometer from yesterday and draw a straight line forward. But the real world is messy. Sometimes a sudden cold front blows in, or a massive storm rolls through, changing everything your simple guess predicted. This is the heart of time series forecasting: the science of predicting future numbers based on past patterns. For decades, computers have been great at spotting straight lines and repeating cycles in data, like the daily rush of electricity or the monthly sales of a store. But they often struggle when the story changes because of outside factors, like a holiday, a new policy, or a sudden shift in the wind.

Recently, scientists have started using Large Language Models (LLMs)—the same kind of super-smart AI that can write stories or chat with you—to help with these predictions. The idea is that if an AI can understand language, maybe it can also understand the "story" behind the numbers, like why a storm might ruin a forecast. However, there's a catch: these AI models are sometimes too chatty or too confident. They might try to "hallucinate" a number instead of doing the hard math, or they might ignore the actual data in favor of a cool-sounding story. The big question is: How do we get an AI to be both a creative storyteller and a strict, math-focused accountant at the same time?

This is where a new framework called CastFSR comes in. Think of it as a three-step team meeting for predicting the future. Instead of letting the AI just guess a number, CastFSR forces it to work in three distinct modes: Fast, Slow, and Reflective.

First, there's Fast-Thinking Forecasting. Imagine this as the "quick sketch" phase. The AI looks at the recent history of the data (like the last few days of wind speed) and quickly picks a specialized, lightweight math tool—like a standard calculator or a simple graphing program—to draw a basic prediction. It doesn't try to be clever here; it just establishes a solid, data-driven "prior" or baseline. It's like a chef tasting a soup and saying, "Okay, this is the base flavor."

Next comes Slow Deliberative Reasoning. This is the "deep dive" phase. Now, the AI stops and thinks hard. It looks at the "story" around the data: Is there a holiday coming up? Did the weather report change? Is there a special event? It retrieves this extra context and asks, "Does this story change the soup's flavor?" If the context suggests a big change (like a storm is coming), the AI carefully adjusts the initial sketch. It doesn't throw the sketch away; it just tweaks it based on the new evidence. It's like the chef realizing, "Oh, it's going to be a spicy night, so I need to add a pinch more chili," but only adding it where it makes sense.

Finally, there's Reflective Evaluation. This is the "quality control" phase. Before the AI gives its final answer, it acts like a strict editor. It checks the prediction against the rules of the real world. For example, if the prediction is for wind power, the AI checks: "Can wind power be negative? No, that's impossible." If the prediction breaks a rule or doesn't match the story it just told, it fixes the mistake. It ensures the final number is not just a guess, but a fact that makes sense in the real world.

The paper shows that this three-step process works really well. When the researchers tested CastFSR on real-world data—like electricity usage, wind power, and stock prices—it beat many other top methods. They found that by separating the "fast math" from the "slow thinking" and adding a "reflection" step, the AI became much more reliable. Even better, they showed that this system can work with huge, expensive AI models right out of the box, but they also taught a smaller, cheaper AI model how to do the same thing by training it to follow these three steps. The result is a forecasting system that doesn't just crunch numbers; it understands the context, respects the rules, and knows when to double-check its work.