Zero-shot forecasting with LLMs

Problem

LLMs can forecast without task-specific training. I tested how reliable general-purpose LLMs are on IoT sensor data from the SensBee smart-city platform.

Approach

• FastAPI microservice for next-value prediction on SensBee.
• 6-stage data pipeline.
• 4 LLMs in 192 experiments.
• Method and results in a written final report.

Results

• 192 experiments, 4 LLMs.
• General-purpose LLMs unreliable on numeric sensor data.
• Grade 1.3.