The paper investigates the effectiveness of large language models (LLMs) in time series forecasting. Through a series of ablation studies on three popular LLM-based methods, the authors find that removing the LLM component or replacing it with simpler attention layers does not degrade performance and often improves it. Despite their computational cost, pre-trained LLMs do not outperform models trained from scratch and fail to capture sequential dependencies in time series data. The study also shows that simpler models, such as those using patching and attention structures, can achieve similar performance to LLM-based methods. Additionally, LLMs do not assist in few-shot learning scenarios. The findings suggest that LLMs are not necessary for time series forecasting and that simpler models may be more efficient and effective. The authors conclude by highlighting the need for further research into the specific tasks where LLMs can be beneficial, such as time series reasoning and social understanding.The paper investigates the effectiveness of large language models (LLMs) in time series forecasting. Through a series of ablation studies on three popular LLM-based methods, the authors find that removing the LLM component or replacing it with simpler attention layers does not degrade performance and often improves it. Despite their computational cost, pre-trained LLMs do not outperform models trained from scratch and fail to capture sequential dependencies in time series data. The study also shows that simpler models, such as those using patching and attention structures, can achieve similar performance to LLM-based methods. Additionally, LLMs do not assist in few-shot learning scenarios. The findings suggest that LLMs are not necessary for time series forecasting and that simpler models may be more efficient and effective. The authors conclude by highlighting the need for further research into the specific tasks where LLMs can be beneficial, such as time series reasoning and social understanding.