Researchers at the Gaidar Institute’s Industrial Organization and Infrastructure Economics Department have discovered that modern language models can forecast changes in industrial production by analyzing news, company statements, and regulatory announcements. This new approach allows for a more accurate assessment of the future state of the economy amid instability and enables a faster response to changes than traditional forecasting methods.
Today, most models for forecasting industrial production rely on official statistics, which are published with a delay. At the same time, economic changes—from sanctions and logistics disruptions to fluctuations in demand—begin to affect the economy much earlier and are reflected almost immediately in the news stream.
The study’s authors set out to test whether large language models can extract useful economic information from news and use it to forecast industrial production indices.
The researchers compared several approaches. The first was based on a classic analysis of news sentiment using the RuBERT model. The second utilized the large language model Llama 3.1, which assessed the impact of news on industry while taking context into account. The third approach proved to be the most sophisticated—a combination of the DeepSeek language model and RAG technology, in which the AI analyzed not only individual news stories but also related materials, including media publications and press releases from the Bank of Russia.
In most industries, news indices significantly improved the quality of output forecasts: forecasts more accurately captured the actual rate of change in production—the average forecast error decreased to 4.41 percentage points.
The greatest effect was observed in sectors closely integrated into the global economy and dependent on external supplies: machinery and equipment manufacturing, metallurgy, and the production of rubber and plastic products. For these sectors, sanctions, logistical disruptions, and supply chain breaks are quickly reflected in the information landscape and provide early signals of changes in output.
News related to output dynamics had the weakest impact in industries with low output flexibility—such as petroleum products and chemical manufacturing—where no significant improvement in forecasts was observed. The reason lies in the long production cycle and high inertia in output: real-time signals from external market conditions do not immediately translate into changes in output volumes due to infrastructure and contractual constraints.
A telling example is the blockade of the Strait of Hormuz: the price shock was reflected in the news almost instantly, but oil companies were physically unable to react just as quickly to the change in supply conditions; therefore, the news signal would not have improved the forecast for short-term output dynamics.
According to the authors, the main finding of the study is that modern language models allow us to move beyond a simple analysis of the emotional tone of news reports to a deeper modeling of the cause-and-effect relationships between events.
“Large language models are beginning to function not as conventional systems for analyzing text tone, but as tools for the economic interpretation of information. The approach involving expanded context proved particularly important: the model analyzes not a single news item, but the entire related economic situation surrounding the event. This allows for a more accurate assessment of the impact of events and more timely forecasting of changes in output dynamics,” notes one of the study’s authors, Mikhail Anikutin, Junior Researcher at the Gaidar Institute’s Industrial Organization and Infrastructure Economics Department.
The authors believe that such technologies could become an important part of economic monitoring systems. For the government, this offers an opportunity to respond more quickly to changes in the macroeconomic context, while for businesses and analysts, it provides an additional tool for assessing the current economic situation and future risks.