MPC Transcripts Reveal Shifting Inflation and Growth Concerns in 2015–17

Bank of England policy discussions from 2015 to 2017 reveal more than the committee’s recorded decisions. An analysis of 258 member statements finds that policymakers’ views on inflation and economic growth shifted over the period, and that those views were linked to their positions on Bank Rate. Perceptions of both above-target inflation and above-potential growth were associated with a greater likelihood of favouring a rate increase. For rate cuts, growth assessments mattered more: weaker views on growth were linked to a greater likelihood of favouring a cut, while the inflation measure had little explanatory power.

The analysis uses recently published transcripts of 29 Monetary Policy Committee (MPC) meetings. The records cover the period from 2015 to 2017, the latest year currently available, and are released with an eight-year lag. The researchers use large language models to turn the discussion into structured measures of individual members’ economic assessments and the arguments behind their policy positions.

Measuring what members said

The data set contains statements from 14 committee members: seven internal and seven external members across changing tenures. The committee has nine seats, though three meetings in 2017 included eight members. The researchers focus on each speaker’s assessment of UK inflation relative to the Bank’s 2% CPI target and of output growth relative to potential, both over the year ahead.

For each statement, a language model assigns separate scores to inflation and growth on a scale from -5 to +5. A score of -5 means the speaker sees the measure as far below its benchmark; zero means close to the benchmark; and +5 means far above. The model used is OpenAI gpt-5-mini, with each score produced from a separate instruction and constrained to a single integer. The measures are intended to capture how strongly members discuss inflation and output stabilisation, rather than merely repeat published economic data.

The researchers also extract the reasoning expressed in each statement. They recast arguments as short causal claims and classify them by topic, type and tense. Each argument can belong to more than one of 14 topic groups, including inflation, labour market conditions, output, exchange rates and policy decisions. The analysis draws on earlier work examining central bank transcripts, including Prat et al (2018). It focuses on six topics for discussion.

Figure 1: Extracting causal arguments and topics from MPC transcripts

Growth concerns gave way to a stronger reading of resilience

The text-based growth measure weakened in the months before the June 2016 referendum and fell sharply around it, before recovering through 2017. That trajectory reflects both movements in realised GDP growth and the uncertainty that followed the result. Financial markets reacted sharply, business surveys deteriorated, and contemporaneous forecasts warned of a marked slowdown and significant recession risks. MPC members discussed the market reaction in their June and July 2016 meetings.

As the economy proved more resilient than initially expected, members’ assessments of growth improved. The recovery in the text measure was partly consistent with lagged GDP data, so it should be read as a record of policymakers’ expressed views rather than a real-time measure of the economy itself. The researchers also find that the relationship between perceived growth and inflation became more positive and statistically stronger after the referendum period. They interpret this as consistent with a greater emphasis on demand-side forces in members’ assessments.

Inflation sentiment moved differently. Members’ assessments of inflation relative to target closely tracked CPI inflation over the sample, with a reported correlation of 0.75. Concerns about above-target inflation increased as inflation rose. This comparison gives the extracted measure an economic anchor, while also showing that transcript-based analysis can add information about how individual policymakers interpreted the data.

Chart 1: Trends in text-based inflation and growth metrics over 2015–17

Discussion topics changed after mid-2016

Across the statements, output was the most discussed topic, accounting for 19% of statements. Inflation represented 15%, while direct references to the policy-rate decision made up 14%. That last category is deliberately narrow: it covers explicit discussion of the rate decision. Other arguments are classified by their immediate subject, even when they could influence the decision indirectly.

After June 2016, the discussion placed greater weight on rate setting, trade-offs between inflation and output, and political events. References to inflation and financial markets declined. The shift suggests that members’ deliberations responded not only to changing economic assessments but also to a changed mix of immediate concerns. Topic counts, however, describe what members discussed; on their own, they do not show how much weight a member gave each argument.

Chart 2: Distribution of topic shares in MPC transcripts before and after June 2016

Economic assessments and Bank Rate positions

To test whether the language-based measures help explain policy choices, the researchers examine decisions on Bank Rate, excluding asset purchase decisions. They compare members’ inflation and growth assessments with whether each member favoured an increase, a hold or a cut. The analysis controls for member-specific differences, annual CPI inflation, annual GDP growth, and the Bank’s year-ahead forecasts for inflation and growth. The question is whether expressed views explain differences beyond those aggregate indicators and forecasts.

Across the full sample, stronger perceptions of above-target inflation and above-potential growth were each associated with a higher likelihood of favouring a rate increase. Both relationships were statistically significant at the 5% level. For a hold, stronger inflation concerns were linked to a lower likelihood, while the growth measure showed no significant relationship. For a cut, the pattern differed: stronger growth assessments reduced the likelihood of favouring a cut, while the inflation estimate was small and statistically insignificant.

Chart 3: Bank Rate positions and views on inflation and growth

The member-by-member results broadly echo the pooled analysis. For rate increases, the inflation and growth coefficients were positive for essentially every member, with strong and statistically significant relationships for several. For cuts, growth coefficients were negative across the board: weaker growth assessments were associated with a greater likelihood of favouring a cut. Inflation coefficients for cuts were smaller and varied around zero.

These patterns were not driven solely by a few members. The individual estimates exclude members who took part in fewer than 10 meetings and those who never favoured an increase or cut, so the comparisons cover a selected group. Even with that qualification, the results suggest a common link between members’ stated assessments and their policy positions, while the strength of that link differed among them.

What the transcripts add

Recorded decisions alone can conceal different reasoning: members may reach the same position while weighing inflation, growth and uncertainty differently. The transcript measures offer one way to recover part of that variation and follow how it changed over time. They also show why rate increases and cuts should not be treated as mirror images in this sample: both inflation and growth assessments helped explain positions favouring increases, while growth assessments were the clearer correlate of positions favouring cuts.

The method depends on how models interpret natural language, and the researchers present the scores as structured readings of members’ statements. The results therefore complement the economic data and recorded decisions rather than replacing them. A next step would be to compare the reasoning in these historical transcripts with members’ public speeches from the same periods, testing how policy arguments appear across different forms of communication.

The authors are Dragos Gorduza and Livia Silva Paranhos of the Bank’s Advanced Analytics Division, and Ivan Yotzov of its Structural Economics Division. Related research on transcript analysis includes Bennani and Romelli (2024), Aruoba and Drechsel (2026), and research on economic uncertainty.

Readers can contact the authors at bankunderground@bankofengland.co.uk or leave a comment through the comments page.