2608.07251-Reading Copom's Tone: A Weighted LLM Framework for Hawkish-Dovish Sentiment, Forward Guidance, and Uncertainty

Introduction to the show: ident: Paper Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Next we'll be talking about the paper "Reading Copom's Tone: A Weighted LLM Framework for Hawkish-Dovish Sentiment, Forward Guidance, and Uncertainty".

Jane: The paper was written by Gabriel de Macedo Santos from Instituto de Tecnologia e Liderança.

Tom: Stay tuned as we take you through the paper and discuss its implications.

Title: Tom: Welcome back. Today's paper comes from Gabriel de Macedo Santos at the Instituto de Tecnologia e Liderança, and it's about reading the tone of Brazil's central bank. The Copom, which is the committee that sets Brazilian interest rates. This author built a system that tries to measure whether the bank's official statements sound hawkish or dovish using a large language model.

Jane: For anyone who doesn't live in monetary policy world, hawkish means leaning toward higher rates to fight inflation, and dovish means leaning toward easier money. That's the entire game in a nutshell. And the interesting part is that the model reads the actual Portuguese text of the statements, sentence by sentence, rather than just the headline decision.

Lu: The clever part, I think, is that they don't just ask the model for one overall label. They break each statement into sentences, classify each sentence, and then also pull out short phrases inside those sentences with an intensity weight. So you get a sense of not just how many sentences lean one way, but how strongly the language is phrased.

Meng: Which matters because central banks are famously careful. They rarely say "we are definitely raising rates next meeting." They say things like "the balance of risks remains unfavorable," and you need to know whether that sentence is carrying real weight or just routine filler.

Tom: Exactly. And the author is upfront that this is inspired by a system called iSent from Itaú, one of Brazil's big banks. But the paper extends that idea in several directions, which we'll get into shortly.

Jane: The sample runs from August 2016 through August 2026 and covers eighty statements. That's a long window that includes some very different economic eras in Brazil — a deep recession recovery, a major inflation spike, and the recent period where inflation stayed stubbornly high.

Lu: And the whole thing is built incrementally, so you can process a new statement the day it comes out without re-annotating the entire history. That makes it usable as a monitoring tool rather than just a one-off research project.

Meng: Right, and I want to know what it actually found over that decade. The numbers in the paper could tell us whether Brazilian central bank communication has had a persistent hawkish tilt or whether it swings with the cycle.

Tom: That's exactly where the summary comes in. It has real numbers on the sentence mix and the historical path of the tone index.

Jane: Let's go there.

Summary: Tom: So the results. The paper classified 1,498 sentences across those eighty statements. Neutral language turns out to be the biggest category at 42 point 1 percent. Hawkish sentences are 33 point 3 percent, dovish 18 percent, and about 6 point 5 percent are what they call out of context — administrative material and procedural sentences that shouldn't count toward tone at all.

Jane: That distribution makes sense. Central bank statements spend a lot of words describing projections and decision mechanics. If every sentence were scored as pushing a direction, that would mean the bank is shouting, and they never shout.

Lu: And the weighted score — the author's addition to the basic iSent idea — comes out positive on average. The mean document score across the sample is plus 0 point 107, so a mild hawkish tilt over the full decade. The most hawkish reading is August 2021 at plus 0 point 570, right in the middle of the tightening cycle when Brazilian inflation was running hot.

Meng: The most dovish reading is January 2017 at negative 0 point 357, during the easing cycle. And the paper groups the window into regimes: dovish on average through 2020, then sharply hawkish in 2021 to 2023 at plus 0 point 263, and still positive in 2024 through 2026 — even though the guidance score in that latest stretch is near zero.

Tom: That last point is really the heart of the paper. Tone and forward guidance are correlated — the contemporaneous Pearson correlation is 0 point 719 — but they're not the same thing. You can have a statement that sounds hawkish in its diagnosis of inflation risks while the actual guidance about the next rate move is ambiguous.

Jane: And the paper has a separate structural layer just for that. It scores guidance direction, how explicit the guidance is, uncertainty level, and whether uncertainty rose or fell relative to the prior meeting. That's the innovation I keep coming back to.

Lu: Right, because a pure sentiment score can't tell you whether the committee is signaling a hike or just complaining about the economy. Separating those dimensions is genuinely useful for anyone who reads these statements for a living.

Meng: The author is honest that this is descriptive, not predictive. No backtest against actual Selic decisions or market returns yet. But as a monitoring index, the numbers look economically coherent.

Tom: Coherent is the right word. And the paper itself lays out a detailed roadmap for turning this into something more rigorous. That's what I want to dig into now — what the author says needs to happen next.

Jane: Good, because that's where the real work is. A descriptive index is one thing; a validated tool is another.

Improvements: Tom: So the improvements the paper recommends. The biggest one is a human-labeled benchmark. Right now the same model that reads the text also supplies the classification and the intensity weights, and there's no economist-labeled holdout set to check against. The author wants precision, recall, F1 scores, a confusion matrix, and agreement statistics across human annotators.

Jane: That's the standard the serious central bank text literature has been moving toward. Without that gold set, you can't tell whether the model is actually good or just consistent with itself. And the author flags stochastic instability too — running the same prompt multiple times can give different labels, so you'd want label agreement and score dispersion across repeated runs.

Lu: Exactly. And model dependence. If you swap the underlying LLM or change the prompt, do the extreme readings stay stable? The paper recommends rank correlations across alternative model and prompt settings. That kind of robustness testing is what would make the index credible over time.

Meng: There's also a methodological improvement around aggregation. The current score takes the average intensity of all hawkish signals in a document and multiplies it by the hawkish sentence count. The author admits this isn't the same as summing sentence-level weighted probabilities. A strongly hawkish sentence and a weakly hawkish sentence get the same count weight, so they suggest reporting an unweighted benchmark and a sentence-level weighted sum as alternatives.

Tom: And there's a technical fix as well. The sentence splitter is a simple regular expression, and missing spaces after punctuation can merge sentences that should be separate. That matters because if one merged segment contains both hawkish and dovish phrases, it only gets one class and the mixed signal is lost. A proper Portuguese tokenizer would fix a lot of that.

Jane: Then the validation program extends to economics. The paper wants a study of whether the tone score lines up with actual Selic decisions — contemporaneous and one meeting ahead — and a market event study on DI futures, the Brazilian interest rate derivatives. With transaction costs, volatility scaling, and an out-of-sample split.

Lu: What I appreciate is the implementation audit in the appendix. The author distinguishes what's operational from what's still a research extension. For instance, the FAISS retrieval infrastructure exists but isn't actually used in classification. That kind of transparency is rare.

Meng: It makes the whole thing auditable down to the sentence level. Every score can be traced back to the specific expressions and weights the model extracted. For a central bank monitoring tool, that traceability is arguably the most valuable feature.

Tom: Right, and it sets up the deeper question of what the author was trying to accomplish in the first place. Let's go back to the opening pages, because the abstract and introduction frame the whole contribution.

Jane: And they also give us the latest reading — the August 2026 statement — which is the perfect case study for why this layered approach matters.

First Page: Tom: Going back to the opening page then. The paper starts from a strong claim — central bank communication is not ancillary, it shapes expectations about the reaction function and the path of interest rates. That's the Blinder line of thinking, and there's evidence that forward guidance language can matter more to markets than descriptions of current conditions.

Jane: The problem the paper identifies is measurement. A Copom statement contains simultaneous signals about inflation, activity, exchange rates, fiscal policy, external risks, and expectations. A single sentence might describe weakening activity while warning that inflation expectations are unanchored. A pure word count misses that context, and a document-level label hides disagreement between sentences.

Lu: So the unit of analysis is the sentence, and the design separates rhetorical tone from the structural layers of guidance and uncertainty. The three contributions the author lists are documenting the exact scoring formula, presenting a consistently defined sample from August 2016 onward, and providing an implementation audit that says what is operational versus what still needs validation.

Meng: The latest reading illustrates why that separation matters. The August 5, 2026 statement scores plus 0 point 232 — net hawkish — with eight hawkish sentences, two dovish, and nine neutral. The hawkish content is about upside risks to inflation, unanchored expectations, a resilient labor market, and exchange rate depreciation risk. But the guidance direction is zero, ambiguous, with explicitness at 0 point 5, meaning conditional.

Tom: And on top of that, uncertainty is at level three, the highest, with a change of plus one from the prior meeting. So the economic interpretation is not simply "hawkish statement." It's a hawkish risk diagnosis with no clear commitment on the next move, and rising uncertainty. That's a much richer read than any single sentiment label could give you.

Jane: It's exactly the kind of situation where economists could disagree about what the bank is signaling. The diagnosis points one way, the guidance stays conditional, and the uncertainty makes everything fuzzier. A system that separates those dimensions at least forces you to be explicit about which one you're reacting to.

Lu: And that's the point the author makes about the strong correlation of 0 point 719 between tone and guidance. It's strong but far from perfect, and there's plenty of room for divergence. The August 2026 statement is the proof.

Meng: The title of the paper really captures the distinction. Reading the tone of the committee is not the same as decoding its policy path. The tool is designed around that separation.

Tom: So that brings us to where the paper ends. The conclusion is honest about what this system can and can't do. Should we wrap this up?

Jane: Let's.

Conclusion: Tom: So to wrap up — the paper delivers a monitoring tool, not a crystal ball. It documents a weighted sentiment framework for Brazilian central bank statements, built on sentence-level classification with intensity weights, plus a separate layer for guidance direction, explicitness, and uncertainty.

Jane: And the key finding over eighty statements is a mild hawkish tilt, with an average score of plus 0 point 107. The regime shifts are clear: dovish in the late twenty-teens, sharply hawkish during the 2021 to 2023 inflation fight, still positive in the recent period. And the latest statement shows exactly why a single label fails — hawkish in diagnosis, ambiguous in guidance, high and rising uncertainty.

Lu: The author is explicit that this is descriptive. It doesn't predict Selic decisions or market returns. But the architecture is transparent, incremental, and auditable down to individual sentences and the expressions the model extracted. That's the real contribution.

Meng: And the roadmap for validation is clear — human-labeled benchmarks, model-version controls, better tokenization, aggregation alternatives, and a genuinely out-of-sample test against interest rate decisions and DI futures. It'll be exciting to see that work, especially if it can be benchmarked against the correlations Itaú published for iSent.

Tom: I think the paper's biggest value is that it separates the hawkishness of the diagnosis from the direction of policy guidance. Even if the scores change with future model versions, that separation is a useful discipline for anyone reading central bank communication.

Jane: Agreed. And with that, we're done with this one. Gabriel de Macedo Santos should be proud of the clarity here. We'll be back shortly for the next paper.

Tom: See you soon.