📈 economics

Are Whitepaper Claims Reflected in Market Structure? A Contamination-Aware Pipeline and a Power-Limited Null

This paper presents a contamination-aware pipeline that reveals a previously reported link between cryptocurrency whitepaper narratives and market structure was an artifact of data errors, while an honest null result on the cleaned dataset indicates that current text-based methods lack the statistical power to detect weak alignment, though they can confidently rule out strong correspondence.

Original authors: Murad Farzulla

Published 2026-09-23

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Murad Farzulla

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the world of finance, there is a long-standing debate about how much a company's story matters to its stock price. On one side, traditional economic theory suggests that if a company publishes a clear, useful plan, the market quickly absorbs that information and prices the asset accordingly. On the other side, a field known as narrative economics argues that the stories we tell ourselves about money can sometimes detach from reality, driving prices based on hype or emotion rather than actual utility. Cryptocurrency offers a unique laboratory to test this tension. Unlike traditional stocks, which reveal their true nature slowly through years of earnings reports and analyst reviews, crypto projects often lay out their entire purpose, technology, and goals in a single document called a whitepaper at the very moment they are born. This raises a compelling question: do the functional promises made in these founding documents actually show up in how the tokens behave in the market? If a project claims to be a tool for fast payments, does its token trade like a payment system? If it claims to be a secure vault for wealth, does it act like a store of value?

A researcher named Murad Farzulla set out to answer this by building a rigorous test to see if the words in forty-three cryptocurrency whitepapers matched the actual market behavior of those tokens. The study focused on a specific period from January 2023 to December 2024, gathering hourly price data for dozens of assets. The researcher first used advanced language software to read the whitepapers and categorize them into ten functional themes, such as "store of value," "smart contracts," or "data storage." These categories formed a map of what each project claimed to be. Next, the researcher calculated seven different statistical measures for each token's market performance, such as how volatile its price was, how much it traded, and its overall trend. The goal was to see if the map of claims lined up with the map of market behavior. Before running the final test, the team had to clean their data carefully. They discovered that their initial collection of documents was contaminated with errors: about a quarter of the files were either broken downloads or, in a few striking cases, the wrong documents entirely, such as a file labeled "Cosmos" that was actually the whitepaper for a different network called Binance Smart Chain. Once these errors were removed and the documents were verified to be genuine, the analysis began.

The results of the study were surprisingly quiet. When the researchers aligned the map of claims with the map of market behavior, they found no significant connection. The patterns in the whitepapers did not predict how the tokens moved in the market. To put this in perspective, if the market were a perfect reflection of the whitepaper promises, the two maps would have matched closely; instead, they were essentially unrelated. The researchers were careful to explain that this does not prove the stories are irrelevant, but rather that their specific test could not find a link. The study was limited by the small number of projects available and the difficulty of using language software to accurately capture the complex, technical nuances of crypto documents. The software used to read the whitepapers was not perfectly reliable, which made it harder to detect even a weak connection if one existed. The study was powerful enough to rule out a strong, obvious link between the text and the market, but it could not definitively say whether a subtle, hidden connection exists or if the two are completely separate.

A particularly instructive part of the research involved a false alarm that occurred during the data cleaning process. In the initial, unverified version of the data, the researchers thought they saw a clear pattern: specialized tokens with narrow uses seemed to align with their market behavior, while broad infrastructure tokens did not. However, this pattern vanished completely once the data was cleaned. The apparent pattern had been an illusion created by the contaminated files; the "specialized" tokens that seemed to align were actually based on the wrong documents. This finding serves as a cautionary tale for future research, showing how easily a spurious result can be manufactured if the source documents are not verified. The final conclusion is that while the whitepapers of these forty-three projects are rich with claims, those claims do not appear to be reflected in the structure of the market during the studied period. The market seems to be driven by other forces, or the connection is too faint for current methods to detect, leaving the relationship between crypto narratives and market reality an open question.

Technical Summary: "Are Whitepaper Claims Reflected in Market Structure? A Contamination-Aware Pipeline and a Power-Limited Null"

Problem Statement
The paper investigates whether the functional narratives articulated in cryptocurrency whitepapers correspond to the actual market behavior of their tokens. While efficient-market theory suggests informative narratives are quickly impounded into prices, narrative economics posits that stories can decouple prices from fundamentals. Prior research has largely focused on predicting issuance success or post-listing returns based on whitepaper quality. This study shifts the focus to testing the contemporaneous structural correspondence between two representational spaces: the semantic claims of whitepapers and the cross-sectional market statistics of the tokens.

Methodology
The author develops a three-stage, contamination-aware pipeline to measure structural alignment:

NLP Claims Extraction:
- Data: 43 content-verified whitepapers (spanning assets like BTC, ETH, ADA, etc.) collected from official sources. A critical step involved screening for "corpus contamination," where an earlier version of the dataset contained failed-download stubs and wrong-document files (e.g., a Cosmos entry containing Binance Smart Chain text).
- Classification: Whitepaper text is segmented into 500-word chunks and classified into 10 semantic categories (e.g., Store of Value, DeFi, Privacy) using zero-shot classification via BART-large-MNLI.
- Reliability Check: The study explicitly measures inter-method reliability, comparing BART-NLI against DeBERTa-v3 and a local LLM. The mean pairwise correlation ( $\rho$ ) is approximately 0.31, with some pairs showing near-zero agreement, identifying the text instrument as a binding constraint on statistical power.
Market Statistics Construction:
- Data: Hourly OHLCV data from Binance for 49 assets (Jan 2023–Dec 2024), yielding 17,543 timestamps.
- Features: Seven cross-sectional financial metrics are computed for each asset: mean return, volatility, Sharpe ratio, maximum drawdown, average volume, vol-of-vol, and price trend.
Alignment and Inference:
- Alignment: The claims matrix ( $C \in \mathbb{R}^{43 \times 10}$ ) and market statistics matrix ( $S \in \mathbb{R}^{43 \times 7}$ ) are aligned using Orthogonal Procrustes rotation to find the optimal rotation minimizing the Frobenius norm.
- Metric: Structural similarity is measured using Tucker's congruence coefficient ( $\phi$ ), which equals cosine similarity without mean-centering.
- Dimension Handling: Two estimators are used: a "zero-padded" variant (padding the 7D market matrix to 10D, which mechanically caps the maximum possible $\phi$ at 0.70) and a "matched-dimension" variant (reducing the 10D claims matrix to 7D via SVD), which allows for a theoretical $\phi$ of 1.0.
- Power Analysis: A positive-control Monte Carlo simulation injects known congruence into the real market factor matrix to determine the minimum detectable effect (MDE). The study finds that due to the low reliability of the text instrument ( $\rho \approx 0.31$ ), the realistic MDE for 80% power is $\phi \approx 0.66$ .

Key Results

Corpus Contamination Artifact:
- An initial analysis of a contaminated corpus ( $n=37$ ) suggested a structural ordering where "specialized" tokens (e.g., privacy, DeFi) aligned more strongly with market structure than "infrastructure" tokens.
- Upon content verification and removal of stubs/wrong documents, this ordering collapsed. In the clean corpus ( $n=43$ ), no single token registers as significantly helping or hurting alignment, and the apparent specialized-vs-infrastructure split disappears. This is reported as a contamination diagnosis rather than a finding.
Power-Limited Null Hypothesis:
- Primary Estimate: The matched-dimension alignment yields $\phi = 0.303$ ( $p=0.353$ ). The conservative zero-padded estimate is $\phi = 0.223$ ( $p=0.459$ ).
- Interpretation: Both estimates are well below the moderate similarity threshold of 0.65.
- Power Constraints: The positive-control simulation confirms the pipeline is sensitive (recovering injected signals accurately). However, the low reliability of the NLP instrument means the study has over 80% power only to detect strong alignment ( $\phi \ge 0.70$ ). It lacks the power to distinguish between weak alignment ( $\phi \approx 0.3$ ) and no alignment.
- Conclusion: The result is an "absence of evidence" for strong alignment, not "evidence of absence." The data are inconsistent with strong narrative-market coupling but cannot rule out weak coupling.
Robustness:
- The null result holds across rolling time windows, subsample resampling, jackknife tests, and alternative alignment metrics (RV coefficient, distance correlation, CCA, PLS).
- Bitcoin, despite its market dominance, has negligible impact on the alignment metric ( $\Delta \phi = -0.001$ ).

Significance and Contributions
The paper frames its contribution as a methodological framework and a cautionary tale for text-based narrative economics:

Methodological Contribution: It introduces a reproducible, contamination-aware pipeline for comparing textual and market representational spaces, coupled with a positive-control power apparatus that explicitly bounds what such comparisons can detect.
Empirical Finding: It reports a power-limited non-detection of claims-market alignment on a clean corpus. The author emphasizes that while strong alignment is ruled out, the study cannot adjudicate between weak alignment and no alignment due to instrument noise.
Cautionary Tale: The study demonstrates how undetected corpus contamination (e.g., wrong-document files) can manufacture spurious cross-sectional signals. It argues that content-level verification (word counts, provenance checks) is a necessary precondition for text-market analysis.
Limitations: The author explicitly states that the findings do not license claims that whitepaper analysis offers "limited value" or that narratives are secondary. The results are limited by sample size ( $n=43$ ), the low reliability of off-the-shelf zero-shot classifiers in the crypto domain, and the temporal mismatch between founding-era whitepapers and 2023–2024 market data.

The paper concludes that the question of whether whitepaper claims reflect market structure remains open, requiring larger corpora, domain-adapted classifiers, and time-stamped narrative data to move beyond the current "bounded null."

Technical Summary: "Are Whitepaper Claims Reflected in Market Structure? A Contamination-Aware Pipeline and a Power-Limited Null"

More like this