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

This paper documents an applied natural-language-processing framework for measuring the tone of Brazilian Monetary Policy Committee (Copom) statements. The author motivates the work by noting that "Central bank communication is not ancillary to monetary policy. It shapes expectations about the reaction function, the likely path of short-term interest rates, the balance of risks, and the conditions under which policymakers may change course. The practical difficulty addressed is that Copom statements contain simultaneous signals about inflation, activity, the exchange rate, fiscal policy, external risks, expectations, and the degree of confidence in the baseline scenario, and A pure word count can miss this context, while a document-level LLM label can hide internal disagreement across sentences."

The project is explicitly inspired by iSent, Itaú's Central Bank sentiment classifier, particularly its sentence-level division of official communication into hawkish, dovish, neutral, and out-of-context classes. The paper notes that iSent classifies sentences from official central-bank documents as dovish, neutral, hawkish, or out of context, and aggregates the relative presence of those classes into an index between-1 and 1. The authors adopt that intuitive unit of analysis and class taxonomy, while deliberately changing the calibration and output architecture. The paper emphasizes that The project is independently developed and is not affiliated with or endorsed by Itaú Unibanco.

The implementation extends iSent's idea in three directions. First, an LLM identifies short hawkish and dovish expressions and assigns each a 0-to-1 intensity weight. Second, the document index combines sentence counts with document-specific average signal intensities, producing a bounded score from -1 to 1. Third, a separate full-document layer measures forward-guidance direction, guidance explicitness, uncertainty level, and change in uncertainty.

The empirical sample covers 80 Copom statement dates from August 31, 2016 through August 5, 2026. The August 2016 cutoff is deliberately applied before computing any descriptive statistic, sentence share, correlation, extreme reading, or regime average in this paper. The sample contains 1,498 classified sentence objects within the retained window, of which 1,400 enter the score denominator; 98 out-of-context sentences are excluded. The primary source is the official Portuguese text of Copom statements.

Text normalization and segmentation: "New lines are replaced with spaces before classification. Sentence boundaries are identified with a regular expression that looks for whitespace following periods, question marks, exclamation points, semicolons, or colons when the next token begins with an uppercase letter, number, quotation mark, or parenthesis. Segments with fewer than four whitespace-delimited tokens are excluded. The authors acknowledge this deterministic splitter... inherits formatting defects from scraped source text, as Several archived annotations contain missing spaces after punctuation, causing logically separate sentences to be merged."

Sentence classification: Each segment is sent to an LLM under a Portuguese prompt that defines four mutually exclusive classes. Hawkish sentences emphasize persistent or elevated inflation, upside risks, restrictive policy, higher rates, fiscal deterioration, or other forces that increase the required degree of monetary restraint. Dovish sentences emphasize disinflation, economic slack, weak activity, easing, or prospective rate cuts. The prompt also asks the model to extract supporting expressions with weights: 0.1-0.3 for weak, 0.4-0.6 for moderate, and 0.7-1.0 for strong evidence.

Decision-only rule: "A special instruction treats a sentence that only reports the mechanical rate decision as neutral with no signals. This is economically motivated: the index is intended to capture the communication surrounding the action, not mechanically encode the rate change twice."

Historical anchors: The prompt includes three historical calibration anchors: March 2020 as strongly dovish, May 2021 as hawkish, and August 2023 as dovish. These anchors guide the LLM; they are not hard constraints on the final score.

Weighted document score formula: For statement t, with N H, N D, N N the counts of hawkish, dovish, and neutral sentences, and w-bar H, w-bar D the average intensities of all extracted hawkish and dovish signals in the document:

Score t = (w̄ H × N H,t − w̄ D × N D,t) / (N H + N D + N N)

The score is then clipped to the interval [-1, 1]. Positive values indicate a net hawkish tone and negative values a net dovish tone. Neutral sentences dilute the magnitude without changing the numerator. A subtle implementation detail: intensity averages are computed from all extracted signals, not only from signals found in sentences whose final class matches the signal.

Structural layer: "Guidance direction equals-1 when the text points to easing, 0 when the next move is unclear, and +1 when it points to tightening. Guidance explicitness equals 0 for no guidance, 0.5 for indirect or conditional guidance, and 1 for explicit guidance. Uncertainty level ranges from 0 to 3, while uncertainty change takes values-1, 0, or +1 relative to the prior meeting. The composite is GuidanceScore t = Direction t × Explicitness t, which is not part of the tone index."

Incremental execution: Existing annotations and structural classifications are loaded from JSON caches, and only dates not already present are sent to the LLM. The complete archive is then rescored so that all outputs remain synchronized.

Corpus composition: "The August 2016-August 2026 sample contains 1,498 classified segments. Neutral sentences are the largest class, at 631 observations or 42.1% of the corpus. Hawkish sentences total 499 (33.3%), dovish sentences 270 (18.0%), and out-of-context sentences 98 (6.5%). The LLM extracted 1,552 hawkish phrase signals and 787 dovish signals within the retained sample," with mean weights of 0.654 and 0.586 respectively.

Descriptive statistics: The average document score is +0.107, with a median of +0.106, standard deviation of 0.208, and interquartile range of -0.041 to +0.265. Negative / nonnegative scores are 26 / 54.

Historical regimes: "From August 2016 through 2020, communication is dovish on average (-0.078), consistent with repeated disinflation language, slack, and easing guidance. The series turns sharply positive in 2021-2023, with an average of +0.263... The 2024-August 2026 window remains strongly positive at +0.237, even though its average guidance score is near zero." Mean uncertainty levels rise across eras from 2.23 to 2.54 to 2.76.

Extreme readings: The most dovish observation in the retained sample is January 11, 2017, at -0.357, and The strongest hawkish reading is August 4, 2021, at +0.570, supported by fourteen hawkish sentences and no dovish sentences. The August 2021 statement emphasizes persistent inflation, an unfavorable composition of prices, elevated fiscal risk, an upside-skewed balance of risks, and an explicit expectation of another adjustment of the same magnitude.

Tone and guidance correlation: The contemporaneous Pearson correlation between the tone score and guidance direction multiplied by explicitness is 0.719; the Spearman correlation is 0.718. This is a strong but incomplete association. The authors caution: Because both outputs are generated by LLM prompts applied to the same documents, this correlation is descriptive and should not be interpreted as independent validation.

Latest reading (August 5, 2026): The latest statement scores +0.232, with eight hawkish, two dovish, and nine neutral sentences. Hawkish content "includes an upside-skewed balance of inflation risks, unanchored expectations, a resilient labor market, services inflation risk, exchange-rate depreciation risk, and the possibility that demand stimulus weakens monetary transmission. The structural layer assigns guidance direction 0 and explicitness 0.5, meaning the statement contains conditional guidance but does not clearly commit to either continued cuts or renewed tightening. Uncertainty is level 3 and change +1. The economic interpretation: It is a hawkish risk diagnosis combined with conditional, directionally ambiguous next-step guidance."

The paper identifies four principal limitations. First, "the absence of a manually labeled holdout set. The same LLM that interprets the text supplies both the class and the signal weight. Without economist annotations, confusion matrices, calibration curves, or agreement statistics, the model's accuracy cannot be quantified. Second, segmentation issues: Missing spaces after punctuation can merge independent clauses or sentences. Because one merged segment receives one class, the output can understate mixed communication. Third, intensity aggregation concerns: The score multiplies a sentence count by the mean weight of every same-sign signal in the document... It gives identical count weight to a strongly hawkish sentence and a weakly hawkish sentence while scaling the entire hawkish count by the document's mean signal intensity. Fourth, the project does not yet test whether the score predicts Selic changes, DI returns, yield-curve shifts, or forecast errors."

The paper also notes that strict end-to-end replication would also require the raw statements, model identifier, prompt version, and inference settings, as The raw statements file and external model configuration are not included in the supplied artifact set.

"The paper makes three contributions. First, it documents the exact scoring formula and the data lineage from sentence annotations to document scores. Second, it presents a consistently defined historical sample from August 2016 to August 2026, excluding all earlier observations from the empirical analysis. Third, it provides an implementation audit that separates what is currently operational from what remains a validation or research extension."

The authors are explicit about scope: The main contribution is therefore methodological: a transparent, incremental, auditable system that separates rhetorical tone from policy guidance and uncertainty. They conclude: "The project's most credible current contribution is measurement and organization, not prediction. It provides a disciplined way to separate the tone of the economic diagnosis from the direction and explicitness of future policy guidance. Converting this monitoring system into a publishable predictive index requires a human-labeled benchmark, model-version controls, segmentation improvements, aggregation ablations, and a genuinely out-of-sample Selic and DI validation exercise. They also clarify that The index is a tone measure, not a probability. A score of +0.23 does not mean a 23% probability of a rate increase, and that Cross-period comparisons should consider document length and source formatting, making the score... most reliable as one input to a broader communication dashboard that also reports sentence counts, guidance, uncertainty, and the underlying excerpts."

Improvements derived from the paper:

1. Replace the regex sentence splitter with an LLM-based segmenter that recognizes sentence boundaries even when punctuation spacing is missing, and that merges/keeps clauses according to economic meaning rather than typographic tokens. The improved system will correctly separate merged sentences from scraped text, reducing understatement of mixed hawkish/dovish communication.

2. Add a human-labeled holdout set with economist annotations for sentence class and signal intensity, plus confusion matrices and inter-annotator agreement metrics. The improved system will report calibrated accuracy, precision/recall per class, and a quantitative confidence band for every document score.

3. Abolish the current intensity aggregation (multiplying sentence count by document-level mean signal weight) by instead summing per-signal weights directly, or by applying a frequency-weighted geometric mean. The improved system will assign equal contribution to each extracted expression, so a document with one strong signal and many weak signals no longer behaves identically to one with uniformly moderate signals.

4. Add a dedicated mixed-communication detector that flags documents where hawkish and dovish sentences co-occur with high intensity on both sides, and outputs a disagreement score rather than a single net value. The improved system will expose internal tension—e.g., a hawkish risk paragraph next to a dovish guidance paragraph—instead of collapsing it into one number.

5. Implement model-version control and prompt-version pinning for the exact LLM identifier, temperature, and prompt hash, and store these in the output JSON cache. The improved system will enable exact replication, auditability, and reproducibility across future model updates, with automatic re-scoring only when the prompt or model changes.

6. Introduce dynamic historical anchors—instead of hard-coded March 2020, May 2021, and August 2023—that are selected per subsample period or per policy cycle. The improved system will adapt calibration to changing monetary regimes while preserving intertemporal comparability through a meta-anchor invariant.

7. Add a stand-alone forward-guidance parser that does not use the same LLM call as tone classification, using a separate, more structured prompt that outputs JSON fields for direction, conditionality, state-contingency, and horizon. The improved system will produce guidance scores that are only weakly correlated with tone by construction, enabling a cleaner decomposition of communication channels.

8. Implement an out-of-sample predictive validation module that tests the tone and guidance scores against Selic changes, DI futures returns, yield-curve shifts, and analyst forecast errors, using a rolling window with a burn-in period. The improved system will convert the monitoring index into a decision-support tool with measured predictive value, not just descriptive correlation.

9. Add a document-length and formatting normalization layer that controls for the number of sentences, the presence of bullet points, and the ratio of out-of-context segments. The improved system will report a length-adjusted score and a confidence interval, making comparisons across statements robust to source formatting differences.

10. Create a multi-signal, per-sentence intensity recalibration where the LLM also outputs a confidence level for each extracted expression, and where weights are recalibrated via a small economic validation set (e.g., market reactions around statement releases). The improved system will learn soft intensity values from market data rather than relying on arbitrary 0.1–0.3/0.4–0.6/0.7–1.0 buckets.

11. Add a temporal dynamics module that tracks the change in tone score and guidance score from meeting to meeting, including the uncertainty-change field, to detect shifts in policy communication before actual rate changes. The improved system will provide an early-warning signal for Copom turning points.

12. Expose explainable sentence-level and phrase-level evidence alongside every score so users can inspect exactly which expressions drove the hawkish/dovish classification, the intensity weights, and the structural guidance labels. The improved system will produce a full audit trail with quoted Portuguese text and English glosses, enabling analysts to verify or override any reading.

What the improved AI system can do:

• Generate a fully reproducible, auditable, and human-verifiable central-bank communication dashboard with separate metrics for tone, intensity, guidance direction, explicitness, uncertainty level, and uncertainty change.

• Correctly parse malformed scraped text and segment mixed statements without losing the signal of merged sentences.

• Produce calibrated confidence intervals and accuracy statistics that are honest about the model's limitations.

• Distinguish between the rhetorical tone of the economic diagnosis and the policy intention, allowing users to see when a hawkish-sounding statement still leaves the next move ambiguous.

• Alert analysts to internal contradictions (e.g., hawkish inflation talk vs. dovish guidance) and track how those contradictions evolve across meetings.

• Predict the direction and magnitude of future Selic changes and DI yield-curve movements with quantified, out-of-sample performance metrics.

• Support a what-if workflow: an analyst can edit a sentence or phrase and see the resulting change to the tone score and guidance label, enabling stress-testing of communication strategies.

• Monitor changes in uncertainty over time and correlate them with economic variables, yielding insight into how Copom's communication evolves during tightening and easing cycles.