thought 2

* Desiderata

 

Looking back to when I was working at the NIMS, all the research that NIMS conducted was aimed at supporting decision-making in the field.

 

Accurate weather forecasting technology is directly linked to safety and survival throughout the industry and the country, and it's impact is huge.

 

For example, when a wind warning is issued, ships are unable to operate and many inquiries and complaints are pouring in. Because it is directly related to their survival.

 

Therefore, highly accurate forecasting technology is really important, and additionally, it also need to be explainable.

 

 In other words, forecasting models must be reliable to ensure that their forecasts are well-founded.

 

 Furthermore, forecasting models must be adaptable to new forecasting environments, so it can make forecasts with high accuracy in a variety of environments.

 

Considering these things, time series foundation models with multimodal data is promising that can be applied to this field. 

 

I think LLM's few-shot, zero-shot capability might allows us to respond quickly to situations that require quick decision-making such as typhoons and earthquakes, and to flexibly cope with constantly changing situations such as climate change.

 

* Research direction

 

Recent researches on timeseries foundation model using LLMs are really interesting and surprising.

 

But I found several things to improve. (This is just my opinion!) 

 

First, representations of time series which is used to align with text needs to be "more delicate".

 

Second, most of the timeseries forecasting models deal with univariate. They didn't deal with multivariate simultaneously and not consider inter-correlation between variables. Some ideas have come up in this regard, but implementation is likely to take some time.