The jury is out on LLM for time series forecasting

 

Large Language Models have all been the rage in data science and has been an area of focused attention by many in finance, yet it is less clear whether these models provide added value versus other techniques for forecasting time series. Now we have some evidence on their viability and the results are not positive. See "Are Language Models Actually Useful for Time Series Forecasting?" This is not an easy paper to read because it requires a fair amount of key knowledge about ML and LLM models, yet the extensive analysis is strong evidence that LLM may not be a solution to time series forecasting. 

The researchers test times series data across a wide spectrum and run ablation tests that find that dropping the LLM component will improve the quality of forecasts. Obviously, different researchers will generate different results based on their implementation of the techniques, but this study places the burden on the LLM community to show why LLM will be better than simpler and less costly approaches to forecasts.