NYU Researchers Train GPT-3 to Make Surprisingly Accurate Time Series Predictions by Breaking Numbers into Digits

NYU researchers trained GPT-3 to predict next events in time series data represented as numeric digit strings.

They re-tokenized the numbers so each digit was separated, allowing better sequence predictions.

The resulting LLMTime model matched or exceeded specialized time series models without tuning.

LLMTime works by preferring simple rules to complete sequences, a form of Occam's razor.

While good at predicting, LLMTime sometimes made incorrect deductions about time series behavior.