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

Technical Summary: "Are Whitepaper Claims Reflected in Market Structure?"

Problem Statement
The paper investigates whether the functional narratives articulated in cryptocurrency whitepapers correspond to the actual structural behavior of their tokens in financial markets. 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 whitepaper quality as a predictor of initial coin offering (ICO) success or post-listing returns. This study shifts the focus to contemporaneous structural correspondence: do the semantic categories of a project's founding claims align with the cross-sectional market statistics (return, risk, liquidity) observed over a two-year period (2023–2024)?

Methodology
The authors develop a content-verified, contamination-aware pipeline to measure structural alignment between two representational spaces:

1. Narrative Space (Claims Matrix $C$): Whitepapers for 43 assets were content-verified to remove failed-download stubs and wrong-document files (e.g., a "Cosmos" entry that was actually Binance Smart Chain text). Text was 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. The resulting scores were aggregated into a composition matrix $C \in \mathbb{R}^{43 \times 10}$.
2. Market Space (Statistics Matrix $S$): Hourly OHLCV data (17,543 timestamps) were used to compute seven cross-sectional financial statistics for each asset: mean return, volatility, Sharpe ratio, maximum drawdown, average volume, vol-of-vol, and price trend. These were z-normalized to form matrix $S \in \mathbb{R}^{43 \times 7}$.
3. Alignment and Inference: The two matrices were aligned using Orthogonal Procrustes rotation. Congruence was measured using Tucker's congruence coefficient ($\phi$).
   • Dimension Handling: Because the spaces differ in dimension (10 vs. 7), the authors report two estimates: a zero-padded estimate (conservative lower bound) and a matched-dimension estimate (SVD-reduced, no padding).
   • Power Analysis: A positive-control Monte Carlo simulation injected known congruence into the real data to determine the minimum detectable effect. The analysis highlighted that the binding constraint is the reliability of the text instrument (inter-method agreement $\rho \approx 0.57$ within categories), not the estimator itself.

Key Results

1. Contamination Diagnosis: An earlier version of the corpus suggested a structural signal where "specialised" tokens aligned more strongly than "infrastructure" tokens. The authors demonstrate this was entirely an artefact of corpus contamination. Once documents were verified and re-collected, this ordering vanished; no token registered as significantly helping alignment on the clean corpus.
2. Power-Limited Null: On the clean sample of $n=43$, no significant claims–market alignment was detected.
   • Matched-dimension estimate: $\phi = 0.303$ ($p = 0.365$).
   • Zero-padded estimate: $\phi = 0.223$ ($p = 0.433$).
   • Both are well below the authors' conservative benchmark for moderate similarity ($\phi \ge 0.65$).
3. Power Constraints: The study is underpowered to distinguish weak alignment from none. The realistic minimum detectable effect (at 80% power) is $\phi \approx 0.44$. The design has >80% power to detect strong alignment ($\phi \ge 0.70$) but cannot adjudicate between weak alignment ($\phi \approx 0.3$) and no alignment.
4. Positive Control: A simulation confirmed the pipeline is sensitive; it recovered injected cross-domain congruence almost exactly (e.g., injected $\phi=0.80$ recovered as $0.806$). This confirms the non-detection reflects limited power and instrument reliability, not an insensitive pipeline.
5. Robustness: The null result holds across rolling time windows, subsample resampling, jackknife analysis, and alternative alignment metrics (RV coefficient, distance correlation, CCA, PLS).

Significance and Claims
The paper frames its contribution as a methodological advance and a cautionary tale rather than a definitive proof of narrative-market decoupling.

• Methodological Contribution: The authors introduce a reproducible pipeline for comparing textual and market representational spaces that explicitly accounts for corpus contamination and bounds detectability via power analysis. They argue that content-level verification (word counts, provenance checks) is a precondition for text–market analysis, as undetected contamination can manufacture spurious cross-sectional signals.
• Interpretation of the Null: The authors explicitly state this is "absence of evidence for alignment, not evidence of its absence." The results are inconsistent with strong alignment ($\phi \ge 0.70$) but cannot rule out weak alignment.
• Limitations: The authors highlight three binding limitations:
  1. Power: The sample size ($n=43$) and low-reliability text instrument prevent distinguishing weak effects from zero.
  2. Construct Validity: The zero-shot classifier shows modest inter-model agreement and fails a face-validity check (e.g., it does not correctly identify Bitcoin's primary "Store of Value" narrative, ranking it sixth). The instrument may not capture price-relevant narrative nuances.
  3. Temporal Mismatch: The corpus consists mostly of founding-era documents (some pre-2018), while market data is from 2023–2024. This static-vs-contemporaneous mismatch likely attenuates any true alignment.

The paper concludes that while the data is inconsistent with strong alignment between whitepaper claims and market structure, the design cannot adjudicate among the possibilities of a genuine disconnect, an under-powered design, or a low-validity text instrument. Future work requires larger corpora, domain-adapted classifiers, and time-stamped narrative sources.