Mind the Shift: Decoding Monetary Policy Stance from FOMC Statements with Large Language Models

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

Tom: I'm Tom.

Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.

Tom: Today's paper: "Mind the Shift".

Jane: The gist The Delta-Consistent Scoring (DCS) framework is an annotation-free method that maps frozen large language model representations to continuous hawkish–dovish stance scores by jointly modeling absolute stance and relative…

Tom: First, who's behind it and why it matters.

Title and authors: Tom: So we're looking at this paper, "Mind the Shift: Decoding Monetary Policy Stance from FOMC Statements with Large Language Models," and what it's doing is trying to figure out the hawkish or dovish stance of central bank communication. It seems like they're tackling a real problem because just labeling every statement individually doesn't capture how policy actually works in the real world.

Jane: Right, that's the core issue they point out, and it’s about relativity. The paper suggests that treating each statement as isolated is missing something important because market reactions depend on how the tone shifts between different meetings >

Lu: Exactly. They argue that stance isn't just about a single statement; it’s relative to what came before and what comes next, which is a key idea for representation learning in this area >

Meng: From my side, I'm interested in how they handle the training since we often deal with models that are frozen. They introduce this Delta-Consistent Scoring, or DCS, which aims to map those frozen representations into continuous scores by looking at both the absolute stance and the shift between meetings >

Lalam: The paper says DCS learns an absolute stance score for each statement and a relative shift score between consecutive statements using a dual-axis projection module >

Tom: That sounds like they're trying to bridge the gap between what’s happening in one statement and the sequence of statements that make up a meeting >

Jane: And their training objective is really clever, it uses this delta-consistency loss to make sure that changes in those absolute scores line up with the estimated relative shifts >

Lu: The math looks like Ldelta equals E

z abs t − z abs t−one − alpha · tanh(z rel t / tau τ): , which is forcing the absolute score changes to match the relative shift estimates > <ref:2603.14313#pg1>

Meng: I'm looking at that loss function, and then they add an auxiliary confidence regularizer, Lconf to minimize the entropy of those absolute stance predictions >

Title and authors: Tom: That regularization part is interesting because it helps with calibration, making sure those continuous scores are actually well-scaled rather than just arbitrary numbers >

Jane: And that leads into their evaluation where they show DCS consistently beats supervised probes and LLM-as-judge baselines on sentence-level classification, getting up to seventy-one point one percent accuracy across four different models >

Lu: That's a solid result, especially when you look at the performance across such a range of model scales, from one billion parameters up to fourteen billion parameters >

Meng: But it’s not just about accuracy on text; they show economic relevance too, correlating the meeting-level scores with inflation indicators and Treasury yield movements >

Tom: That's where I get really interested. If these scores actually track real economic things like CPI changes, that gives the whole thing a lot more weight than just being good at classifying words >

Jane: And they validate this by showing strong correlations with year-over-year changes in CPI and PPI, with the best result coming from DeepSeek-R1-DistillQwen-14B reaching a Spearman correlation of zero point six two three seven for inflation >

Lu: What's really compelling is that DCS maintains positive correlations with both CPI and PPI across all four policy periods they tested, under both Pearson and Spearman metrics >

Meng: That temporal robustness over different policy regimes, P1 through P4, suggests the framework isn't just overfitting to one specific economic cycle or tone of language >

Tom: I see what they mean here: the frozen LLM representations are encoding information that is truly actionable for macroeconomic analysis rather than just linguistic noise >

Jane: And to test that, they looked at an ablation study where removing Ldelta caused the biggest drop in accuracy and CPI correlation, proving that the temporal structure modeling is the main learning signal >

Lu: That confirms that without enforcing those relative shifts, you just don't get a good representation of how policy moves over time >

Meng: And removing Lconf also hurt the CPI correlation significantly, which tells us calibration of those absolute scores is just as important as capturing the temporal structure itself >

Title and authors: Tom: So it sounds like they’ve found a way to systematically extract an economically meaningful and continuous measure from FOMC communication without needing any manual stance labels >

Jane: It gives practitioners a systematic way to quantify that hawkish or dovish stance embedded in the text for use in financial decision-making >

Lu: The paper concludes that DCS successfully decodes monetary-policy signals by aligning absolute scores with relative shifts between consecutive meetings >

Meng: For deployment, it means you can take these continuous scores and plug them directly into your economic models to see how they react to policy changes >

Tom: So, the big picture here is that we can move away from just looking at what a single sentence says and start looking at the entire trajectory of communication across multiple meetings >

Jane: It's a practical tool for anyone trying to analyze central bank statements for real-world economic impact, and it's definitely something worth exploring further >

Lu: I think the strength of this work is really showing that the internal representations in these large language models are already rich with monetary policy information that just needs the right way to decode it >

Meng: For production, as long as we can get a frozen LLM representation and feed it through this DCS pipeline, we have a systematic method to extract these signals >

Tom: So that's what they did with "Mind the Shift: Decoding Monetary Policy Stance from FOMC Statements with Large Language Models," mapping those LLM representations to continuous stance scores by modeling both absolute stance and relative inter-meeting shifts. We've seen how it performs, how it validates its economic relevance, and why the temporal structure is so crucial for capturing policy signals >

Jane: That’s what we have on this paper today, showing that even without manual labels, you can extract continuous monetary policy stance scores that actually align with inflation and Treasury yields.

The paper's summary: Tom: So we’re diving deeper into this paper now, "Mind the Shift," and what it really does is take those massive frozen language models, like Llama or DeepSeek, and turn them into a continuous score for how hawkish or dovish they are about monetary policy.

Jane: Right, so it’s not just classifying one statement in isolation anymore. This system looks at the whole sequence of statements from a meeting and tries to figure out the overall stance trajectory by linking the absolute tone of each line with how that tone shifts between meetings >

Lu: The core idea is that stance isn't inherent in a single sentence, it’s relative to what came before and what comes after. They are using this delta-consistency loss to force the model’s internal representations to respect that temporal relationship between statements >

Meng: From a deployment angle, they're doing this without needing any human labels for stance. The training objective is all about making sure the absolute score changes align with the estimated shift between meetings, which is a clever way to self-supervise the temporal structure >

Tom: And they didn't stop there; they added an auxiliary confidence regularizer. That helps keep those absolute scores well-calibrated so you don't get some weird, meaningless numbers for stance >

Jane: The results are pretty impressive on the text classification side, showing that this method consistently beats other methods on sentence-level scoring by about seventy-one point one percent accuracy across four different models >

Lu: But what really makes me interested is the economic validation. They tested these meeting-level scores against real economic data, and they found strong correlations with things like inflation indicators and Treasury yield movements >

Tom: That’s the kicker, right? If it’s just a linguistic trick, it doesn't matter much. But if these continuous scores actually track inflation or market reactions, then this isn't just some academic curiosity >

Jane: And they show robustness across different policy regimes—P1 through P4—meaning the method holds up whether the central bank is in a normal rate cycle or a near-zero rate environment >

Lu: The ablation study confirmed that removing the delta-consistency loss crushed both accuracy and the correlation with inflation, which proves that modeling those relative shifts is actually what gives you those real-world signals >

Meng: So for us in production, this means we can take a frozen model, run it through this DCS pipeline, and get a continuous measure of policy stance without any manual annotation effort >

Tom: It’s systematic. It’s continuous. It moves beyond just reading the text and starts decoding the movement of policy signals across time itself >

Jane: So, if you're looking at FOMC transcripts for economic analysis, this provides a structured way to extract that hawkish or dovish signal that actually has some connection to how inflation is moving and what the markets are pricing in >

Lu: The implication is that we can now use these AI-derived scores as a more reliable feature in our macroeconomic models because they are grounded in temporal consistency rather than just isolated text features >

Meng: We need to make sure this doesn't break at scale, obviously, but the fact that it works across different model sizes suggests it’s pretty efficient for what it does >

The paper's improvements: Tom: So, beyond just getting a score, the paper is proposing ways to actually improve how you use these scores for analysis >

Jane: They are suggesting that we need to focus on making sure those continuous scores are really useful, not just accurate on a test set >

Lu: The authors point out that the single-axis version of their approach falls below the full model because separating the absolute stance from the relative shift is crucial for both classification and economic validity >

Tom: That makes sense. If you only know if a statement is hawkish or dovish in isolation, you’re missing half the story about how policy actually evolves >

Jane: They show that by keeping both the absolute score and the relative shift score separate, you get better results for both sentence classification and those real-world correlations with inflation >

Lu: The authors also suggest that for practical use, we need to focus on calibration. That’s why they added the auxiliary confidence regularizer, Lconf >

Meng: I see that as something we have to think about when we deploy this. If the absolute scores are just wildly jumping around without a proper scale, the whole economic correlation thing falls apart >

Tom: Exactly. The ablation study showed that removing Lconf significantly hurt the inflation correlation, proving that if you don't calibrate those absolute scores properly, you lose the link to what actually matters economically >

Jane: So they are suggesting a two-part refinement process: first, use the delta-consistency loss to capture the temporal structure between meetings >

Lu: And second, use that confidence regularization term to keep those absolute stance predictions on a sensible scale >

Tom: It’s like training the model with two different goals simultaneously: one for sequencing and one for stability >

Jane: It means that the next version of this system needs to be focused on balancing those two losses so you get that strong economic alignment we saw, where the scores track CPI changes consistently across policy regimes >

Lu: The implication is that this isn't just a single training objective; it’s a way to structure the learning process so the AI learns both linguistic pattern and economic reality at once >

Meng: If you want to push this into production, you need to design your pipeline around these two learned components: the absolute score and the relative shift estimate >

Tom: So, what we see here is that they’ve moved past just getting a high accuracy number; they’re focused on building a system that produces continuous signals with actual economic meaning >

Jane: And what this means for us listening right now is that we can expect future papers in this area to focus heavily on how to make those confidence scores more reliable for real-time financial modeling, because that's where the current gap lies >

Conclusion: Tom: So we're wrapping up on "Mind the Shift: Decoding Monetary Policy Stance from FOMC Statements with Large Language Models," which essentially shows how to map frozen AI representations into a continuous hawkish-dovish score by looking at both absolute tone and how that tone shifts between meetings >

Jane: It’s really about moving beyond just reading text and actually extracting a consistent, measurable policy stance from the way the AI understands those statements over time >

Lu: The core finding is that temporal self-supervision, using that delta-consistency loss to align absolute scores with relative shifts, works because it respects the inherent relativity of monetary policy communication >

Meng: From a production standpoint, it’s a solid approach for getting continuous signals out of LLMs without needing massive amounts of human annotation data every time >

Lalam: I'm confident in this answer. The model confirms that temporal structure is the primary signal for decoding these policy stances across different backbones >

Tom: And the economic validation is what really seals the deal, showing these scores correlate with things like inflation indicators and Treasury yield movements, which is what matters to us as engineers >

Jane: It changes how we look at central bank transcripts because it gives us a continuous measure of stance that actually lines up with real-world economic conditions across different policy eras >

Lu: The limitation they flag is that the model relies on the frozen LLM representations, so if the underlying language understanding isn't strong, you won't get these good signals >

Meng: So, while it’s a great starting point for feature engineering in economic models, we need to be careful about deploying it without fine-tuning for specific policy jargon because that’s where the frozen representation might break down >

Tom: Right. It’s a strong method for extraction, but we still have to worry about the fidelity of the initial representations coming from those frozen models >

Jane: And this kind of work opens up a lot of avenues for how we can build better financial tools that use these AI-derived continuous scores instead of relying on older classification methods >

Lu: Looking ahead, I think the next step is applying this idea to even more complex temporal dependencies, maybe incorporating external economic data into the training objective >

Meng: If we could make it robust enough to handle real-time data streams rather than just static transcripts, that would be a huge win for deployment speed >

Lalam: I see the potential for scaling this up across different financial markets too; the underlying mechanism seems general enough for that application >