RDP 2025-06: An AI-powered Tool for Central Bank Business Liaisons: Quantitative Indicators and On-demand Insights from Firms 6. Conclusion

Central banks rely on timely, accurate, and nuanced economic signals to assess current economic conditions and support effective monetary policy, especially during periods of heightened uncertainty. The RBA's extensive business liaison program has long provided qualitative and quantitative intelligence directly from firms, complementing insights from official statistics and other data as well as econometric models.

This paper introduces a new artificial intelligence-powered text analytics and information retrieval tool to enhance the analysis of the information collected through the liaison program. This system leverages recent advances in NLP to process and analyse around 25 years of confidential liaison information, covering around 22,000 firm-level interactions, with all analytics taking place within the RBA's secure information technology environment. The tool enables analysts to quickly filter, query and aggregate liaison messages at scale, greatly enhancing their efficiency and ability to respond to ad hoc demands and systematically assess information.

It introduces several new capabilities, including rapid document querying, topic and tone classification, quantitative signal extraction (such as firms’ self-reported prices growth), and uncertainty measurement from qualitative text. Our paper also presents some validation exercises showing that the new NLP-based capabilities produce accurate results when manually reviewed against human benchmarks. However, it will always remain critical for economists using the tool to interrogate, validate and apply judgement to any LM-based outputs.

We show that integrating a large number of liaison-based indicators into model-based nowcasts for growth in the official measure of aggregate wages significantly improves nowcasting performance compared to a best-in-class baseline Phillips curve model. Specifically, the best-performing machine learning shrinkage approach (lasso) significantly reduces nowcasting errors by meaningful magnitudes. Our results reinforce the value of liaison intelligence as well as the importance of applying shrinkage-based machine learning methods.

Overall, this work makes several key contributions. It provides the first detailed technical description of a systematised, large-scale text-based analytics and information retrieval system for central bank business liaison data. Our code is also publicly accessible to foster similar implementations and improvements across other central banks. Additionally, the paper extends an emerging literature on NLP applications in macroeconomics and highlights how capturing timely, narrative business information can materially enhance the accuracy and responsiveness of nowcasting models, especially in more uncertain economic conditions.

We consider our paper the tip of the iceberg in terms of integrating transformer-based LMs into information retrieval systems for corporate intelligence. While the LMs we work with in this paper – with 100s of millions of parameters – were considered ‘large language models (LLMs)’ at the time of their release, today a typical LLM has billions or trillions of parameters. With more computing power, future work could use these models to improve on the capabilities we have introduced, as well as introduce new ones, such as tense detection, to construct forward- and backward-looking indicators. It will also be useful to extend our empirical nowcasting exercise to other key macroeconomic variables beyond wages growth to further explore the value of central bank liaison information for predictive models.