When the Market Speaks: A New Dataset That Actually Listens

TL;DR for operators

FinMarBa is a useful reminder that in finance, sentiment is not what a sentence sounds like. Sentiment is what the market does after reading, absorbing, ignoring, overreacting to, or misunderstanding that sentence. Very elegant. Very inconvenient.

The paper introduces a 61,252-headline financial sentiment dataset built from Bloomberg Market Wraps covering 2010 to January 2024.¹ Instead of asking human annotators whether a headline feels positive, negative, or neutral, the authors use a market-based labelling process: extract headlines, identify relevant tickers with GPT-4, observe the next-day price reaction, compare that reaction with the ticker’s rolling five-year return distribution, and assign a label from that relative move.

The operational implication is direct: if the business goal is investment signal generation, then the training label should be tied to market reaction. A model trained to imitate human financial prose judgement may become fluent at describing news. A model trained on market reaction labels has at least some chance of learning what the market historically priced.

The paper’s evidence supports the direction of travel, not a finished trading system. FinMarBaBERT, trained on the new dataset, produces a stronger S&P 500 sentiment signal than FinBERT trained on Financial-Phrasebank: the reported Sharpe ratio is 0.30 versus -0.13 over the 2019–2024 backtest window. The authors also run perturbation tests that shuffle headlines within short windows to examine whether the signal behaves consistently under controlled disruption.

For fintech builders, quant teams, and financial AI product teams, the valuable idea is not “use this dataset and retire rich.” That would be adorable. The valuable idea is: build financial language models around the market response you care about, then validate the resulting signal against costs, execution, risk controls, and regime change.

A headline is not positive because it sounds cheerful

A company reports widening losses. A human annotator reads the sentence and marks it negative. The stock rises the next day.

That is the paper’s opening example, and it is a good one because it exposes the central defect in much financial sentiment work. Human language judgement and market judgement are not the same object. A loss can be “negative” in ordinary language but “positive” to investors if the loss was smaller than expected, if management guided better than feared, if the balance sheet improved, or if the market had already priced in worse news.

Traditional financial sentiment datasets often ask humans to classify text according to perceived economic tone. That is understandable. It is also slightly strange. The downstream use case is frequently market prediction, yet the upstream label is based on human interpretation rather than realised market response.

FinMarBa attacks that mismatch. Its core argument is simple: if financial sentiment classification is used to build market signals, then the label should be informed by the market.

That is the mechanism-first story. The dataset matters. The model comparison matters. The backtest matters. But the real contribution is the replacement of a semantic label with a behavioural one.

The old labelling task asked humans to interpret; FinMarBa asks prices to respond

The paper positions FinMarBa against two common sources of financial sentiment supervision.

The first is Financial-Phrasebank, a widely used dataset of roughly 5,000 financial sentences labelled by humans. Its annotations are useful for teaching models how people classify financial text, but the paper argues that this creates two problems. Human annotators may miss the actual market context, and they may introduce inconsistencies or leakage. More practically, human labelling is slow and expensive.

The second category is social-media-based sentiment data. These datasets may capture investor chatter, but they also include plenty of text from people whose opinions may have little direct influence on asset prices. There is nothing inherently wrong with that if the goal is crowd mood measurement. It is less attractive if the goal is a clean signal tied to market-moving information.

FinMarBa chooses a different source: Bloomberg Market Wraps. The authors collected more than 3,700 daily market wrap reports from 2010 to 2024. These reports summarise major financial events and are already filtered through Bloomberg’s editorial process. From those summaries, GPT-4 extracts concise, market-relevant headlines. The result is not raw news ingestion; it is curated-news condensation.

That distinction matters. FinMarBa is not claiming to capture every piece of information the market sees. It is building labels from a particular institutional information stream: Bloomberg-style daily market summaries. This is a strength for noise reduction and a boundary for generalisation.

The labelling mechanism is the actual product

FinMarBa’s pipeline can be reduced to one operational question:

Which asset should have reacted to this headline, and was its next-day move unusually positive or unusually negative relative to its own history?

The answer is constructed in stages.

The rolling five-year window is important. A one percent move does not mean the same thing for a Treasury yield proxy, an oil future, a megacap equity index, and a volatile single stock. By comparing the next-day move with the asset’s own recent history, the method tries to label market reaction in relative rather than absolute terms.

The sentiment rule is intentionally mechanical. If the next-day move is above the 60th percentile of the historical return distribution, the headline is labelled positive. If it is below the 30th percentile, the headline is labelled negative. Otherwise, it is labelled indecisive.

This is not magic. It is a disciplined proxy. The market does not submit a clean survey response saying, “I rose today specifically because of headline #14.” The method assumes that the relevant ticker’s next-day move is informative about the market reaction to the headline. That assumption is plausible enough to be useful and fragile enough to deserve respect.

The clever part is also the dangerous part

Market-based labelling sounds objective. It is not quite objective.

It removes one form of subjectivity: the human annotator deciding whether the text sounds good or bad. But it introduces other dependencies: Bloomberg’s editorial selection, GPT-4’s headline extraction, GPT-4’s ticker mapping, the choice of next-day horizon, the 30th and 60th percentile thresholds, and the assumption that the observed return is mainly responding to the headline.

That last assumption is the sharp edge. Markets move for many reasons at once. A headline about oil, the dollar, or the S&P 500 can overlap with central-bank communication, macro releases, earnings, geopolitical news, and positioning effects. The paper recognises that headline responsibility for market movement cannot be fully dismissed as a concern. Good. It would be worrying if it did not.

The right way to read FinMarBa is therefore not: “The market has labelled the truth.”

The better reading is: “The authors have built a scalable label that is closer to the target use case than ordinary human semantic annotation.”

That difference is not pedantic. It is the difference between a dataset worth testing and a dataset worshipped by people who enjoy losing money with confidence.

The dataset distribution tells us what changed

After applying the pipeline, the authors obtain 61,252 annotated headlines. They compare FinMarBa with Financial-Phrasebank on label distribution and coverage.

The label distribution differs sharply.

The authors argue that FinMarBa’s distribution is more consistent with market behaviour. Equity markets have historically had a positive drift, so a slightly higher positive share is plausible. More importantly, FinMarBa has far fewer indecisive labels than Financial-Phrasebank.

That may be useful, but it should be interpreted carefully. A lower neutral share does not automatically mean “better”. It means the labelling mechanism is more willing to convert observed moves into directional labels. For model training, that can sharpen supervision. For robustness, it also raises the cost of noisy attribution. If the return move is confounded, the dataset may confidently label the wrong thing.

The paper also contrasts regional and topical coverage. Financial-Phrasebank contains terms that reflect a strong northeastern European focus, including words such as “EUR,” “Finland,” and “Finnish.” FinMarBa’s frequent terms are more aligned with global market drivers, including U.S. stocks, the dollar, and oil. That makes sense given the Bloomberg Market Wrap source and the intended macro-market use case.

This is not merely a cosmetic improvement. A sentiment model trained on narrow regional corporate sentences may learn a different financial language from one trained on global market summaries. In financial NLP, vocabulary is not decoration. It is part of the signal surface.

The backtest is main evidence, not a victory parade

The paper’s main empirical test is straightforward. The authors train a BERT model on FinMarBa and compare it with FinBERT, which was trained on Financial-Phrasebank. They use the same hyperparameters, with the dataset as the key difference. FinMarBa data from 2010 to 2019 is used for training, and the signal is tested on an S&P 500 backtest from 2019 to 2024.

The daily sentiment score is built from the model’s positive and negative headline classifications. In practical terms, the score rises when positive classifications dominate and falls when negative classifications dominate:

The reported result is clear.

The paper interprets this as evidence that market-based annotation produces a better sentiment signal than human-labelled Financial-Phrasebank supervision.

That interpretation is reasonable, with two practical caveats.

First, a Sharpe ratio of 0.30 is not a trading-system miracle. It is a signal-level result. Whether it survives transaction costs, exposure constraints, turnover, latency, capacity, and portfolio integration is a separate question. Many signals look charming in isolation. Then execution costs arrive, and the charm leaves quietly through the back door.

Second, the comparison is about training labels and source data, not only annotation philosophy. FinMarBa differs from Financial-Phrasebank in size, time span, news source, geographic coverage, market relevance, and label construction. The paper’s design tries to keep model training conditions comparable, but the dataset difference is multidimensional. That does not invalidate the result. It does mean the practical lesson is broader than “market labels beat human labels.” It is closer to: market-aligned, larger, globally relevant, professionally curated financial text can produce a more useful signal than smaller, human-labelled sentence collections.

The perturbation tests are robustness checks, not a second thesis

The paper then runs robustness experiments by perturbing headlines within sliding time windows. Headlines are randomly exchanged within windows of 5, 10, and 15 days, with exchange rates from 10% to 50%. The authors examine differences in Sharpe ratios between the FinMarBa-based signal and the Financial-Phrasebank-based signal.

The likely purpose of this test is not to propose a deployable strategy. It is to check whether the signal behaves consistently when the time alignment of information is deliberately disturbed.

The reported Sharpe-difference table remains positive across tested windows and exchange rates. For example, at a 5-day window, the difference rises from 0.50 at a 10% exchange rate to 1.94 at a 50% exchange rate. The 10-day and 15-day windows also show positive differences, though with smaller and less monotonic patterns.

This supports the paper’s claim that FinMarBa’s signal is more robust than the Financial-Phrasebank baseline under their perturbation design. It does not prove that the labels are a clean causal map from headline to market move. Robustness tests can stress a method. They cannot absolve it.

What finance teams should copy is the supervision logic

The most useful business takeaway is not “download FinMarBa and fine-tune something.” That may be useful for research, benchmarking, or a prototype. But the deeper lesson is architectural.

Financial AI systems often fail because the label does not match the business objective. A team says it wants market prediction, but it trains on human sentiment. It says it wants risk intelligence, but it labels documents by broad topic. It says it wants decision support, but it optimises for explanation fluency. Then everyone is shocked when the model becomes an articulate intern with no edge.

FinMarBa points toward a better pattern:

The principle is simple: supervise the model with the consequence you care about.

For investment analytics, the consequence may be return, volatility, drawdown, volume, or cross-asset spillover. For corporate finance, it may be refinancing cost, credit spread, analyst revision, or liquidity pressure. For wealth platforms, it may be client behaviour, redemption risk, or suitability-triggered portfolio changes.

Language is the input. Business consequence should be the label.

Product builders should treat FinMarBa as a benchmark and a warning

For fintech product teams, FinMarBa is attractive because it automates a painful bottleneck. Human-labelled financial datasets are small, expensive, inconsistent, and often misaligned with trading objectives. A market-informed annotation pipeline can scale.

But scaling labels is not the same as scaling truth.

A production-grade version would need controls that the paper does not fully resolve because, frankly, one paper should not be asked to carry an entire trading platform on its back.

At minimum, a business implementation should add:

1. Event disentanglement. A next-day return may reflect multiple overlapping headlines. Production systems should cluster events, control for scheduled macro releases, and separate idiosyncratic from market-wide moves.

2. Abnormal return adjustment. Raw next-day movement is informative, but abnormal returns against sector, factor, or market benchmarks may better isolate headline-specific reaction.

3. Ticker-mapping validation. GPT-4 entity mapping is useful, but errors in ticker assignment directly corrupt labels. Financial entities are messy: subsidiaries, ADRs, indices, futures, ETFs, commodities, and macro proxies do not behave like clean textbook objects.

4. Regime segmentation. A label learned from 2010–2024 includes low-rate years, pandemic shock, inflation shock, and tightening cycles. The same language may price differently across regimes.

5. Execution-aware evaluation. Signal Sharpe is not portfolio Sharpe. Costs, turnover, leverage, capacity, shorting constraints, and timing assumptions decide whether a model is useful outside a notebook.

6. Source diversification. Bloomberg Market Wraps are high quality, but they are one editorial lens. A robust commercial system should test whether the mechanism generalises to filings, earnings transcripts, central-bank speeches, broker notes, social streams, and exchange announcements.

In other words, FinMarBa is an excellent prototype of supervision alignment. It is not a free replacement for research infrastructure.

The real misconception is that prices are clean labels

The tempting slogan is: “Let the market label the data.”

The better slogan is: “Use market reaction as a noisy but relevant supervisory signal.”

That sounds less exciting. It is also less likely to bankrupt someone.

Market prices aggregate information, expectations, liquidity, positioning, forced flows, risk appetite, and occasionally mass delusion wearing a Bloomberg terminal badge. A next-day move can be the market’s response to a headline. It can also be the response to five other things that happened before lunch.

Still, a noisy relevant label can beat a clean irrelevant one. That is the quiet force of this paper.

Human sentiment labels are clean in the wrong coordinate system. Market-based labels are messy in the right coordinate system. For financial NLP, that trade-off may be worth making.

The business value is alignment, not accuracy theatre

The paper should be read as part of a larger shift in AI implementation: the bottleneck is no longer merely model architecture. It is supervision design.

A financial language model trained to sound financially literate is not necessarily useful. A model trained to classify language according to market reaction is closer to the business problem. Not finished. Closer.

For asset managers, this suggests a practical research path: build market-informed datasets for specific asset universes, horizons, and portfolio use cases. For data vendors, it suggests products that provide reaction-labelled text, not just sentiment scores. For wealth platforms, it suggests that “news sentiment” should be calibrated against actual client portfolios and market outcomes, not generic positive/negative labels.

The strongest use case is probably not fully automated trading. It is signal research, risk monitoring, model evaluation, and news triage. Those workflows benefit from labels that are tied to realised market behaviour while still leaving room for human oversight and portfolio context.

The weakest use case is pretending that next-day market reaction is a universal truth oracle. It is not. It is a measurable behavioural response under specific assumptions. Useful, yes. Sacred, no.

Conclusion: the market is a better annotator, but still a flawed one

FinMarBa’s contribution is not that it creates another financial sentiment dataset. The internet has enough datasets, many of them apparently designed to make graduate students question their life choices.

The contribution is that it changes the labelling question.

Instead of asking, “Does this headline sound positive?” it asks, “How did the relevant market move after this headline, relative to its own recent history?”

That question is much closer to how financial sentiment models are actually used. The reported evidence supports the shift: a FinMarBa-trained BERT model produces a stronger S&P 500 sentiment signal than the Financial-Phrasebank-trained FinBERT baseline, and the perturbation tests provide additional robustness evidence.

The boundary is equally clear. Market-informed labels are not pure ground truth. They are structured proxies for market perception. They inherit the limitations of data sources, ticker mapping, time windows, confounding news, and historical regimes.

But for financial AI builders, that is still progress. The market may not tell the whole truth. It does, however, answer the question most human annotators were never asked properly in the first place.

Cognaptus: Automate the Present, Incubate the Future.