Trend Tuesday: đź”® Can AI Predict the Future? A Look at Time Series and LLMs
Art by @basilonmypizza: https://lnkd.in/eF8FkWzN - https://basilhefti.ch/
The turkey thinks the farmer loves him. Every day, it gets fed. Every day, it predicts: “Tomorrow will be another good day.” Until Thanksgiving (at least in some places). That’s the risk of taking things for granted. Recently, the stock market has been moving heavily. Wouldn’t it be great to know what happens next? That’s exactly what time series analysis tries to do.
đź“Š Time series are powerful tools for understanding the world. In space physics, we used them to study solar wind acceleration (see: When the Unexplainable Becomes Explained). During the COVID-19 pandemic, Ryuta and I applied them to case data (linked below).
This was before the era of large language models. In the COVID case, the qualitative signals were striking: clear turning points and trend shifts. The quantitative predictions were less solid.
🧠Niels Bohr once joked: “Prediction is very difficult, especially if it’s about the future.” Forecasting remains a hard problem. Just ask the participants in the Jane Street Real-Time Market Forecasting challenge on Kaggle.
🤖 So, can LLMs help? The answer is still unfolding. A few techniques stand out:
🔹 Temporal Encoders break down time series data into segments or “tokens,” in a similar way as sentences and words are broken into tokens. This opens the door to applying transformer-based architectures directly to temporal patterns.
🔹 Temporal Comprehension Prompts ask LLMs directly. For example: “Day 1, Price 1, Neutral. Day 2, Price 2, Slightly Positive. Day 3, Price 3, Positive. Day 4, Price 4, Very Positive. What is the expected price on Day 5?” It’s surprisingly effective for qualitative forecasting.
🔹 Proximal Policy Optimization (PPO), a reinforcement learning technique, adds another layer. The model makes a prediction, receives a reward if it's accurate, and then updates its strategy, but only within stable bounds. This “proximal” nature prevents wild swings in behavior and encourages gradual, reliable learning over time.
🚀 This is just a glimpse into what’s possible. Time series analysis, and forecasting more broadly, is not only a major AI trend but also a real-world need. Whether in finance, healthcare, or public policy, we face urgent questions about what comes next. And we need all the help we can get.
đź’¬ How are you using forecasting in your work?
- In space physics: https://lnkd.in/e5ZjVB3N
- Swiss covid data: https://lnkd.in/eN89w9MN
Ryuta has many additional interesting posts on time series analysis.
- Kaggle: https://lnkd.in/eCkAwfb6
Follow me on LinkedIn for more content like this.
