Researchers from the Gaidar Institute's Industrial Organization and Infrastructure Economics Department discussed the use of large language models (LLMs) for analyzing economic news. They use AI to determine the semantic tone of texts, which is a key step in building a news index.
As the experts explained, such an index reflects the mood of businesses and consumers and can serve as a leading indicator of business activity—a signal of where the economy is headed even before official statistics are released.
The results of the laboratory's research show that using DeepSeek R1 as a sentiment tagger increases the correlation of the news index with the IPP (industrial production index) by 3–9 percentage points relative to the large language model Llama 3.1, as well as the pre-trained BERT model on a labeled dataset.
Experts have identified several reasons why DeepSeek R1 demonstrates higher efficiency in analyzing economic texts:
- It can reason step by step and take context into account, which allows it to make more meaningful conclusions (for example: “risk of interest rate increase” → negative sentiment, even if the tone is neutral).
- DeepSeek R1 was trained, among other things, on an array of professional economic and technical texts. This allows the model to derive sentiment not at the “positive/negative” level, but in terms of its impact on the economy, i.e., to apply an economic interpretation of events.
- The model uses chains of reasoning, which allows it to interpret ambiguous texts: “On the one hand..., on the other hand...”; “Prices are rising, but demand is falling...”
Thus, a reasoning large language model (e.g., DeepSeek R1) captures the “tone” of the news more accurately than classical approaches, which means that it allows economists to notice changes in economic trends earlier.