|
The most recent Natural Language Processing (NLP) techniques do not just identify predefined terms (keywords), like the earlier-generation would have done.
Rather, the trained language models identify "inflation situations". They look for any news potentially relevant to consumer goods and services prices, based on comparable, but not necessarily identical, historical examples.
The last two years have been a gigantic stress test. Still, our NLP models are detecting inflation items which were not in the training sample.
Our training sample news are from the 1/2018 - 7/2019 period exclusively: no news past that point has been used in our inflation models training (even in the case of our most recently released language models like the Japanese one introduced last summer).
Goods shortages stories were totally anecdotal in the training sample. Yet, the models do catch them pretty well out of sample, a key factor in their ability to signal turning points in the last 2 years. Every single day, we have shortages flagged in the NewsBot (such as eggs in the UK and ski gears in the US, to take just two examples from today's NewsBot).
If we give enough training examples to a well-specified neural network model, it can mimic the thought process of the analyst (economist) scanning through new data points, to a surprising extent.
The only thing is, the machine can do it at a much, much larger scale than an analyst (our models currently analyse 11000 news articles related with inflation on a daily basis, in seven languages).
|
