FRM®: how do central banks use AI for policy and risk?

On this page

A central bank can use artificial intelligence to improve the information behind a policy decision: it can estimate current conditions before official data arrive, examine a wide set of institutions for connected risks, and analyse payment patterns that isolated rules miss. That is different from AI changing the economy itself. For the FRM® candidate, the useful distinction is between an analytical tool used by the bank and a change in productivity, demand or financial behaviour that the bank must respond to.

The question behind the reading

Central banks have several mandates, and each produces an information problem. A monetary policy team needs a view of inflation and activity even though official releases come with delays and revisions. A financial stability team must find concentrations that may be invisible in any single firm's report. A payment overseer needs to see potentially suspicious networks without treating every unusual transfer as conclusive evidence of wrongdoing. The BIS discussion of central banks and AI places data analysis, economic forecasting, financial stability analysis and payment oversight among the applications. The point is to support an accountable decision, not to substitute a model output for a policy mandate.

The distinction matters because a model can be useful without causing a favourable economic outcome. A better estimate of this quarter's activity can help a policy committee avoid reacting to stale information; it cannot create output. A network alert can direct investigators towards a possible illicit flow; it is neither a finding of guilt nor proof that the payment system is now safe. Start an exam answer by naming the task, then the data that the model can combine, then the decision it informs. Only after that should you explain what could go wrong with the tool.

Nowcasting is measurement, not a new source of growth

Nowcasting means estimating the current state of a variable before its official observation is available. A central bank may combine transaction, transport or other frequent data with published series to estimate current activity. The BIS Project Neo describes supervised machine learning applied to more frequent company data for GDP, consumption and inflation estimates. It also says those estimates complement official statistics, which can arrive later and be revised. Project Neo explicitly does not use or train large language models: the project starts with structured data. That qualification prevents a broad answer about generative AI from becoming an inaccurate description of this specific project.

Consider an illustrative example with invented round numbers. An official activity index was last published at 100. During the next reporting interval, a simple model combines timely observations and produces a current estimate of 102. The analyst can compare that estimate with other evidence while waiting for the official release. If the eventual official index is 101, the model made an error of one index point; that observation can be used to evaluate and recalibrate it. Neither the estimate of 102 nor the error means AI added two points of real output. The model has measured a condition, with uncertainty, rather than changed it.

Data speed creates a trade-off. Frequent private data may cover only certain firms or customers; a shift in that sample can look like a shift in the whole economy. A payment series may change when reporting practices change. The published statistic may also be revised after the fact. A sound policy use therefore asks what population the model sees, how the inputs were obtained, whether past errors change across stress periods, and how the estimated result is communicated to the humans making the decision. Fast information is valuable only when its limits remain visible.

Connected financial crime is a network problem

Many payment risks cross institutional boundaries. If each bank examines only its own transactions, it may see a small and apparently ordinary piece of a wider chain. The BIS Project Aurora tested collaborative analytics, machine learning, network analysis and privacy-enhancing technology on simulated cross-institution payment arrangements. The BIS reports improved detection relative to siloed, rules-based methods in that phase, while also identifying legal, data protection, technical and operational obstacles to practical deployment. That is a proof-of-concept finding, not a claim that all live central-bank systems perform the same way.

Use another illustrative invented example. Three institutions each record a transfer of 10 units along a chain. A local rule that flags transfers above 50 units sees nothing unusual at any one institution. An approved shared analysis could reveal that the same value passes through connected accounts in a short sequence and mark the pattern for investigation. The calculation is deliberately simple: each observed transfer is below the local threshold, while the linked pattern may still matter. The correct conclusion is to investigate and validate the alert, not to announce that a crime occurred. A model with more reach still needs an authorised data-sharing basis, privacy controls and a way to manage false positives.

This also helps separate pattern recognition from policy action. A network model may identify unusual structure in payments; the authorities decide what further evidence is required and what powers they may use. A candidate who writes only that AI is faster misses the core issue: the information is fragmented across firms and jurisdictions. A candidate who writes that pooling is automatically permissible misses the governance issue. The best answer connects the detection advantage to the coverage of the data and then notes the safeguards that make its use credible.

AI changes what the bank has to analyse

The same technology can alter the economy that policymakers are measuring. If firms produce more with the same resources, productive capacity may rise. If businesses invest heavily to adopt AI or households expect higher future incomes, spending may rise too. Inflation pressure depends on how supply and demand move relative to one another; there is no reliable rule that AI must always raise or always lower inflation. The BIS also discusses faster algorithmic price changes and altered policy transmission as possibilities. These are channels for analysis, not forecasts that an exam candidate should present as certain outcomes.

For a separate illustrative scenario, suppose an economy could produce 100 units and planned spending was also 100. AI adoption lifts feasible output to 110 units. If planned spending reaches only 104, the extra capacity exceeds the extra spending by six units, suggesting downward pressure on prices relative to the starting position. If planned spending instead reaches 112, spending exceeds the new capacity by two units, suggesting upward pressure. The figures are invented teaching values, not measured GDP, and they do not imply a precise inflation rate. They show why the direction of an inflation answer turns on the relative movements rather than on the word AI alone.

Financial stability adds a separate channel. Institutions using similar models may make similar decisions at similar times. In a shock, that can amplify a common withdrawal of liquidity or sales of assets. New tools can also introduce opaque outputs, data quality problems, privacy concerns and cyber vulnerabilities. The BIS discussion emphasises these governance challenges alongside possible benefits. A policy response should therefore ask both whether the bank's own analytical use is robust and whether private-sector adoption creates new patterns of risk. Those are related questions, but they have different causal chains.

How to structure an exam answer

When you read a vignette, first decide whether AI is the bank's instrument or a force acting on the economy. For an instrument, state the objective, data, model output, human decision and validation concern. For an economic force, state the transmission channel: productivity and supply, investment and demand, pricing behaviour, or correlated financial actions. If both are present, write both chains separately before linking them. This avoids claiming that better nowcasting itself shifts aggregate supply, or that a suspicious-payments model has established wrongdoing.

Name the institution and the maturity of the evidence. Project Neo explores a forecasting tool and describes its boundaries. Aurora's reported performance comes from a simulated collaborative proof of concept. Neither is a universal performance guarantee for every central bank. The FRM program overview helps place the idea in a wider study plan, while the study methodology is a useful place to organise a practice routine around identifying mechanism, assumption and limitation. The exam skill is the causal explanation: identify the information advantage, then the uncertainty that remains.