RDP 2025-06: An AI-powered Tool for Central Bank Business Liaisons: Quantitative Indicators and On-demand Insights from Firms 1. Introduction

Central bank business liaison programs play an important role in shaping monetary policy decisions by providing timely, on-the-ground insights on economic conditions. Since 2001, staff from the Reserve Bank of Australia (RBA) have conducted around 22,000 interviews with a broadly representative sample of firms, industry bodies, government agencies and community organisations (hereafter referred to as ‘firms’) from across the country under the banner of the RBA's liaison program.

This formal program of economic intelligence gathering has been a useful complement to published economic and finance statistics and information gleaned from econometric models in informing the RBA's assessment of economic conditions. For example, on average, since the pandemic there have been around 550 references to liaison in the RBA's policy-related publications. Figure 1 shows that references to liaison in the RBA's Statement on Monetary Policy have grown over time, reflecting that the RBA has increased its reliance on liaison over the past decade. The real-time availability of liaison information, as well as the richness of the qualitative insights it provides, are particularly useful for navigating periods of heightened uncertainty.

As policymakers face more uncertain and complex macroeconomic conditions (e.g. Lagarde 2023; Wolf 2024) it is critical that they can extract signals efficiently and effectively from textual sources of information from firms, such as earnings reports, earnings calls, and information from interviews and focus groups. Analysing these sources of information vastly expands the available information from conventional quantitative data alone, containing information on ‘why’ and ‘how’ firms are responding to unexpected events, decision-makers' expectations, perceived risks and future opportunities (Hassan, Hollander et al 2024). However, this first-hand information is typically subjective, contextual and is not designed and collected for statistical purposes. As such, it can be difficult to quickly systematically synthesise and communicate the high-frequency signals offered by such data in a way that is useful for informing decisions. Reliance on narrative information also exposes decision-making to the risk of unconscious cognitive and behavioural biases.^[1] Fortunately, recent developments in natural language processing (NLP) offer the prospect of capturing the nuance and richness in this type of information in an efficient and more systemic manner. This increases both the breadth and depth of the analysis and can reduce exposure to unconscious cognitive and behavioural biases.^[2]

In this paper, we introduce the RBA's text analytics and retrieval tool for quantitative and qualitative analysis of liaison messages. The new tool was created in late 2022 and is being used to support economic analysis. It uses modern techniques in NLP to improve analysts' abilities to quickly search and extract signals from liaison text about economic conditions. In particular, the tool introduces three new capabilities: (1) the ability to quickly query the entire history of liaison meeting notes spanning around 22,000 liaison meetings over the past 25 years; (2) the ability to zoom in on particular discussion topics (for example, ‘supply chains’) to examine the level of interest in the topic over time and the associated tone and uncertainty of the discussion; and (3) the ability to extract precise numerical values from the text, such as firms' self-reported growth in wages and final prices. The information underpinning the tool is automatically updated daily, drawing in text data from new liaisons once the meeting notes are finalised. Almost instantaneous insights, at the firm level as well as aggregated, can then be accessed using a simple no-code application.^[3]

The first aim of this paper is to offer a detailed overview of the technical architecture underpinning our text analytics and retrieval tool for central bank liaison information and to disseminate the techniques we use to enrich and extract signals from the associated text. In doing so, we hope to contribute to the development and deployment of similar tools across the central banking community; we have engaged with several other central banks already over recent years to support this. To this end, replicable code for the back and front-end of our tool as well as for our nowcasting exercise is publicly available on GitHub (https://github.com/RBA-ER-Replication/Liaison_ai_tool) and in the Supplementary Information accompanying this paper.

We also contribute to the literature on textual analysis and information retrieval systems for macroeconomics. To our knowledge, this paper is the first to apply modern techniques in NLP to develop an analytics and information retrieval tool for confidential and (near) real-time intelligence gathered through central bank business liaison programs. In doing so, we build on a growing body of work by scholars, central banks, think tanks and international organisations using techniques in NLP to draw insights from other text-based sources of corporate information, such as earnings calls, to address macroeconomic questions related to policy as well as for their economic surveillance activities. Recent examples include the use of quarterly earnings calls to analyse the effect of trade exposures on bank lending (Correa et al 2023); firms' price-setting behaviour (Windsor and Zang 2023); firm-level climate exposures (Sautner et al 2023); supply and demand imbalances (Gosselin and Taskin 2023); cyber risk exposures (Jamilov, Rey and Tahoun 2023); political risk exposures (Hassan et al 2019); sources of country risk (Hassan, Schreger et al 2024); and the diffusion of disruptive technologies (Kalyani et al 2021). In its World Economic Outlook, the International Monetary Fund has used text from earnings calls to monitor firm-level inflation expectations (Albrizio, Simon and Dizioli 2023), while the European Central Bank and RBA are using earnings calls to monitor various aspects of firm sentiment (see Andersson, Neves and Nunes (2023) and RBA (2022a)). NLP techniques have also been applied to the Federal Reserve's Beige Book, which is a summary of commentary on economic conditions from a wide range of business and community leaders across the United States, gathered as part of their own liaison and business survey program, released eight times a year. The Beige Book has been shown to contain useful information about inflation trends (Gascon and Werner 2022) and economic sentiment (Filippou et al 2024).

Finally, we make an economic contribution. We explore the benefits of integrating our liaison-based textual indicators into machine learning predictive models, with an application to nowcasting growth in Australia's official wage price index (WPI). Our approach is compared to a best-in-class baseline Phillips curve model that has been used at the RBA. Incorporating liaison-based indicators significantly enhances nowcasting performance for wages growth, increasing accuracy by around 20 per cent over the full sample. In the case of assessing labour market outcomes at least, this validates the ongoing use of liaison to inform the RBA's assessment of economic conditions. Interestingly, the predictive gains appear to be driven by a small number of variables – that is, the signal is sparse. This is an important caveat to the lack of empirical support that has emerged for sparse models in a variety of other predictive applications in macroeconomics (Giannone, Lenza and Primiceri 2021).

The rest of this paper is organised as follows. Section 2 provides background to the RBA's liaison program; Section 3 outlines the technical solution architecture underpinning our tool; Section 4 presents and discusses the new surveillance and analytical capabilities enabled by the tool; Section 5 explores how these capabilities can be used to inform predictive problems by presenting a machine learning exercise based on nowcasting wages growth; and Section 6 concludes.

Footnotes

These include confirmation bias, conservatism bias, interview bias, measurement bias, near-term bias, response bias and sampling bias. While these biases can be pervasive and difficult to detect, the RBA's liaison team has adopted several strategies to minimise them in the liaison process and reduce their influence on qualitative and quantitative research and analysis. [1]

This notwithstanding, while NLP techniques can mitigate some biases, they are not immune to introducing or perpetuating other biases, for example related to language and context. [2]

Meetings conducted by the RBA's liaison staff are undertaken on a confidential basis. This aids in firms being comfortable to share figures and firm-level insights openly with the RBA, as does the trust built over years of engagement with participants. Before being shared more widely across relevant areas of the RBA, information from the liaison program is aggregated, de-identified and summarised. Reflecting the RBA's confidentiality commitments, this information is shared at an industry- or economy-wide level. Similarly, only the RBA's liaison staff, data scientists and select technology support workers have access to the tool. [3]