{"id":"d08a66c9-a3d6-4ce5-844e-a4fc20688111","arxiv_id":"2601.07664","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"After controlling for hidden factors, crypto returns in 2023–2024 are priced by crypto market and size factors plus software-industry, stock-market, profitability, and Fear & Greed factors.","lead":"This paper applies a latent-factor asset-pricing model to weekly returns of 253 large cryptocurrencies in 2023–2024 and estimates which observable risks — crypto market, size, technology-stock, and sentiment factors — carry risk premia after accounting for hidden factors. It is worth reading as an empirical test of whether crypto is becoming integrated with equity markets, though the results are fragile and mostly significant only at the 10% level.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The integration premia (Softw, RS, RMW) are 10%-significant in a 22-factor table with no multiple-testing correction. Under BH-FDR at q=0.10, even the smallest G-X p-value (0.008) fails to reject; the central claim is not yet protected against false positives.","rationale":"The paper's central claim is that crypto returns are priced by equity factors (Software, overall market, profitability). The reported evidence consists of three 10%-level G-X premia in a table of 22 factors. This is a clear multiple-testing setting, and the paper does not correct or even discuss it. The BH calculation is straightforward: with 22 tests and q=0.10, the first BH threshold is 0.00455, so the minimum p=0.008 is not rejected. Therefore the three highlighted factors would certainly not survive. The FM robustness is only partial: Software is strongly significant in FM, so the Software-specific claim is more plausible, but RS and RMW are not robustly significant in FM. Since the abstract/conclusion emphasize all three, the inference is fragile. The EM fill-in concern raised by the reader is legitimate and should be checked, but it is upstream and not directly observable from the reported numbers; the multiplicity issue is directly observable and sufficient to require a more conservative interpretation. Conditional acceptance is still appropriate because the paper is explicitly preliminary and the deficiency can be addressed by reporting FDR-adjusted p-values, running the permutation test, and adding a single-factor robustness check.","tokens_in":7266,"tokens_out":8633,"duration_ms":89413,"concrete_test":"Reapply Benjamini-Hochberg FDR (q=0.10 and q=0.20) to all 22 G-X p-values in Table 5, using the standard BH step-up procedure. In parallel, run a permutation test that fixes the estimated loadings but randomly permutes the time index of each observed factor (or jointly shuffles blocks to preserve dependence) to generate the null distribution of the max |t| across the 22 premia. If Softw, RS, and RMW have BH q-values above 0.20 and their permutation-based max-t p-values are >0.10, the equity-integration conclusion is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Table 5 reports G-X p-values for 22 observed factors. The abstract/conclusion highlight Softw (0.071, p=0.056), RS (0.064, p=0.066), and RMW (-0.033, p=0.068) as evidence of equity-market integration. All three are significant only at 10%; under a global null of zero premia one expects about 2.2 of 22 p-values below 0.10, and the table contains six (SMB_C, Softw, Fear/Greed, RC, RS, RMW). Applying Benjamini-Hochberg at q=0.10 to the 22 p-values yields a first critical value of 0.00455; the smallest G-X p-value is 0.008, so no G-X factor survives the FDR control. At q=0.20 only SMB_C survives. The Fama-MacBeth column does not rescue the broader claim: Softw is significant there (p=0.003), but RS is not (p=0.157) and RMW is only marginal (p=0.056). Thus the 'overall stock market returns and profitability are priced' component of the central claim rests on p-values that are indistinguishable from chance once multiplicity is accounted for. This concern is independent of the EM-fill-in issue: even with perfectly consistent latent factors, the reported inference is not yet established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper estimates risk premia in the cross-section of cryptocurrency returns using the Giglio–Xiu (2021) three-pass latent-factor estimator on a weekly sample of 253 large cryptocurrencies from January 2023 through December 2024. It considers observed factors that include crypto market, size, momentum, and TVL factors; standard Fama–French equity factors and industry portfolios; and three non-tradeable state variables (Fear & Greed, Altcoin Season Index, and hacked value). Results are compared with Fama–MacBeth estimates. The central claim is that crypto expected returns load on both crypto-specific factors and selected equity-industry factors—especially the Software portfolio, the aggregate stock market, and profitability—consistent with partial integration between crypto and equity markets.","tokens_in":7751,"tokens_out":4436,"duration_ms":48667,"significance":"If the central claim survives scrutiny, the paper would provide timely evidence on the evolving integration of cryptocurrency markets with traditional equity markets, an area with mixed prior findings. The use of the Giglio–Xiu latent-factor framework to handle omitted factors in crypto pricing is appropriate and potentially valuable, as is the inclusion of new non-tradeable factors. The paper also makes a useful methodological contribution by applying the three-pass estimator to an unbalanced panel and by using a block bootstrap for inference. However, the strength of the evidence is currently undermined by several internal inconsistencies and by the fragility of the headline results to multiple-testing corrections. The reported Fama–MacBeth comparison appears to contain an error, and the robustness of the latent-factor estimates to the unbalanced-panel imputation is not established. These issues are fixable, but they must be addressed before the empirical conclusions can be considered reliable.","major_comments":[{"comment":"The text states that the Fama–MacBeth estimate of the crypto market risk premium is 0.164% per week (8.5% annualized), but Table 5 reports a Fama–MacBeth RC premium of 0.112. The discrepancy is material: 0.112 weekly annualizes to approximately 5.8%, not 8.5%. Since the paper's contribution includes the claim that the latent-factor approach yields 'materially different premia' relative to Fama–MacBeth, the table and text must be reconciled. If the table is correct, the reported comparison is overstated.","section":"Section 2.1, Table 5"},{"comment":"Table 1 reports that the crypto small-minus-big factor SMB_C has a sample mean of -10.04% per week, while Table 5 reports a Fama–MacBeth risk premium of -0.083% per week for the same factor. For a traded long-short factor, the Fama–MacBeth risk premium should be close to the factor's sample mean. A 120-fold gap is implausible and suggests an error in factor construction, in the Fama–MacBeth implementation, or in the reported statistics. This undermines confidence in the Fama–MacBeth column and in the comparison between the two methodologies.","section":"Table 1 vs. Table 5, SMB_C"},{"comment":"The paper's headline results—Softw (p=0.056), RS (p=0.066), and RMW (p=0.068)—are each significant only at the 10% level in a table of 22 G-X factors. Under the global null of zero premia, about 2.2 p-values below 0.10 are expected; the table contains six. Applying Benjamini-Hochberg at q=0.10 gives a first critical value of 0.1/22=0.00455; the smallest G-X p-value is 0.008, so no G-X factor survives. At q=0.20, only the crypto SMB_C survives. Thus the evidence that equity-industry factors are priced in crypto returns is indistinguishable from chance once multiplicity is accounted for. The authors should report FDR-adjusted p-values or otherwise justify why a multiple-testing correction is unnecessary.","section":"Table 5, multiple testing"},{"comment":"The paper states that latent factors are estimated on an unbalanced panel by 'repeatedly filling in missing weekly returns with values implied by a K-factor structure' until the filled-in matrix stabilizes, and then applying PCA. No convergence criterion, simulation validation, or robustness to the number of latent factors K is reported. Because the Pass 3 risk premia λ_g = Λγ are linear functions of the estimated latent factors, any inconsistency in the imputed factors or in the choice of K can bias all reported G-X premia. The authors should provide either a Monte Carlo validation of their EM-PCA procedure or, at minimum, sensitivity analyses over K (e.g., K=6,8) and over different convergence tolerances. Without this, the latent-factor results are not yet credible.","section":"Section 2, unbalanced-panel estimation"}],"minor_comments":[{"comment":"The sentence listing industry factors has a typographical error: 'Banks, Insur, and F inindustries' should read 'Banks, Insur, and Fin industries.'","section":"Section 1.1"},{"comment":"The Hacks factor is reported with mean and standard deviation 0.00, although the maximum is 0.02. More decimals are needed to convey the scale; the AR(1) residualization may make the units even more opaque.","section":"Table 4"},{"comment":"The profitability factor is labeled 'RMWS' in Table 5 but spelled 'Robust Minus Weak (RM WS)' in the conclusion; the standard abbreviation is RMW. Please use consistent notation throughout.","section":"Conclusion and Table 5"},{"comment":"Only p-values are reported, not standard errors or confidence intervals. Given that the bootstrap procedure is nonstandard (recentered statistic), reporting percentile or bootstrap-t intervals would improve interpretability.","section":"Table 5"},{"comment":"The choice of K=7 is justified only by 'the Bai-Ng Information Criteria.' Please specify which variant of the Bai-Ng criteria was used and how it was implemented on an unbalanced panel with EM imputation, since the standard Bai-Ng theory assumes a balanced panel.","section":"Section 2"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a timely and interesting question, and the G-X framework is a sensible tool. However, the internal inconsistency in the FM RC premium, the implausible SMB_C FM premium, the fragility of the integration result to multiple-testing control, and the lack of validation for the unbalanced-panel imputation are all load-bearing. These are fixable in revision, so I recommend major revision rather than rejection. I would also encourage the editor to ask for a data/code availability statement, as the paper currently provides no replication materials."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper is a straight application of Giglio-Xiu to crypto with a fresh sample (2023–2024) and some genuinely new non-tradeable factors: Fear&Greed, Altseason, and hacked value scaled by market cap. That is a real contribution. The authors are transparent about the preliminary status, use a recognized estimator, and make a good-faith effort to avoid survivorship bias. The TVL result—spanned by latent factors rather than carrying an independent premium—sits nicely with Brigida (2025).\n\nBut the stress-test note is right, and it lands on the load-bearing claim. The abstract and conclusion highlight Softw, RS, and RMW as evidence of equity-market integration. All three are significant only at 10% in the G-X column, and in a 22-factor table six p-values below 0.10 is roughly what you'd expect under a global null. Applying Benjamini-Hochberg at q=0.10, no G-X factor survives—the smallest p-value is 0.008 against a critical value of 0.00455. The FM column rescues only Softw (p=0.003); RS is insignificant and RMW marginal. So the integration story is not yet established.\n\nThere are also internal inconsistencies that need cleaning up. The text says the FM crypto-market premium is 0.164% weekly; Table 5 reports 0.112. More concerning, the SMB_C factor has a sample mean of -10.04% per week while its FM risk premium is -0.083%—a 120-fold gap the paper never explains. For a traded factor, the cross-sectional premium should be close to the average return; if it isn't, the reader needs to know why. The EM fill-in for the unbalanced panel also lacks a convergence criterion and any robustness check on K. Since the entire second and third passes depend on those latent factors, that is a weak link.\n\nThese are fixable. Report FDR-adjusted p-values or at least acknowledge the multiplicity. Reconcile the RC numbers and explain the SMB_C discrepancy. Add sensitivity around K and the block size. The paper deserves a serious referee—the method is appropriate, the data are new, and the core questions are worth asking. But as it stands, the central claim is conditional, and the revision should be substantial.\n\nFor the reading group, it's a decent case study in how multiple testing can sink a headline result, but I wouldn't cite it as evidence yet.","headline":"A legitimate but preliminary Giglio-Xiu application to crypto with new sentiment factors; the integration claim doesn't survive multiple-testing scrutiny.","tokens_in":8170,"tokens_out":2600,"would_cite":false,"duration_ms":28533,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Cryptocurrency expected returns carry significant risk premia tied to hidden technology and profitability factors, evidence that crypto is partially integrated with equity markets.","keywords":["cryptocurrency","risk premia","latent factors","factor pricing","cross-section of returns","asset pricing","market integration","sentiment"],"falsifier":"Re-estimate the three-pass model on the same data with K=6 and K=8, and with alternative missing-data fill-in rules; if the Software, RMW, and Fear & Greed premia change sign or lose significance, the result is an artifact of the factor count. A sharper test: simulate returns from a known factor structure with missing observations and check whether the EM/PCA procedure recovers the true factors—if it does not, the mapped premia inherit that bias.","tokens_in":7170,"feed_emoji":"📈","tokens_out":7564,"duration_ms":71575,"temperature":0.7,"pith_summary":"This paper asks whether cryptocurrency returns are priced only by crypto-specific risks or also by traditional equity risks. Using a three-pass estimator that extracts unobserved (hidden) factors from the cross-section of 253 large coins, it finds that expected crypto returns carry positive premia on the Software equity industry portfolio, on broad stock market returns, and on a stock profitability factor, alongside strong crypto market and size premia. The same factors estimated with a conventional two-pass regression give materially different—often smaller—premia, which the author takes as evidence that omitting latent risks distorts crypto asset pricing. The results matter because they suggest crypto is becoming partially integrated with equity markets rather than remaining a fully segmented asset class, and because they point to sentiment (Fear & Greed) as a priced state variable.","feed_headline":"Crypto risk premia trace to hidden tech and profit factors","feed_subtitle":"A latent-factor study of 253 coins finds crypto is no longer a fully separate asset class.","key_machinery":"The load-bearing tool is a three-pass latent-factor estimator. Pass one estimates K=7 latent common factors from the covariance of the weekly return panel, iteratively filling missing returns under a K-factor structure before applying principal component analysis. Pass two regresses the observed factors (crypto and stock factors plus non-tradable state variables) on the latent factors to obtain a mapping Λ. Pass three estimates latent risk prices γ via a cross-sectional regression, then maps them into observed-factor premia via λ=Λγ. Inference uses a moving-block bootstrap with block length 8 and 1000 resamples.","core_discovery":"On the paper's own terms, the central discovery is that a latent-factor asset-pricing model for weekly cryptocurrency returns over 2023–2024 assigns statistically significant risk premia to equity factors: the Software industry portfolio (0.071% per week), aggregate stock market returns (0.064%), and the stock profitability factor RMW (−0.033%), alongside the crypto market (0.471%) and a strongly negative crypto size premium (−1.345%). Shocks to the Fear & Greed sentiment index also carry a negative premium (−0.051%). The paper interprets these as evidence that crypto and traditional equity markets are increasingly integrated, and that unobserved common factors must be controlled for when es","pith_inferences":["The sample is a short (104-week), mostly rising crypto market; the premia, especially the extreme negative size premium, may be specific to that regime and could dissipate in a longer or bear-market sample.","The three-pass mapping cannot cleanly separate a 'Software' equity factor from a broader technology or risk-on latent factor, so the software premium may be a proxy for a general tech-market channel rather than a distinct industry risk.","A natural out-of-sample check would extend the panel into 2025 (including the post-sample Bybit hack episode) and re-estimate with several alternative latent-factor counts K to test whether the software, profitability, and Fear & Greed premia survive."],"forward_implications":["If the positive Software, stock-market, and profitability premia are real, crypto is partially integrated with equity markets; portfolios of cryptocurrencies incorporate technology and profitability risk.","The large negative small-minus-big premium implies investors demand a premium for holding large-cap crypto, or equivalently a strong preference for larger cryptocurrencies.","The substantial gap between latent-factor and conventional premia (e.g., crypto market 0.471% vs 0.112% weekly) means omitting hidden factors can badly misstate the price of risk.","Priced Fear & Greed shocks suggest sentiment variables belong in crypto asset-pricing models alongside tradeable factors.","TVL's lack of a distinct premium after latent-factor control confirms it is spanned by broader crypto risk rather than an independent priced factor."],"fun_headline_variants":["Latent factors reveal crypto's hidden equity risk premia","Crypto pricing: hidden factors tie coins to tech and profit","Study: crypto risk premia driven by unseen stock factors","Crypto not isolated: latent factors price tech and profit","2023-24 crypto returns: equities lurk in latent space"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The latent factors recovered by iteratively filling missing weekly returns under a seven-factor model must faithfully represent the true unobserved risks; if the fill-in is inconsistent or the number of factors is wrong, every mapped risk premium in the paper is biased.","fun_headline_variants_meta":{"raw":{"variants":["Latent factors reveal crypto's hidden equity risk premia","Crypto pricing: hidden factors tie coins to tech and profit","Study: crypto risk premia driven by unseen stock factors","Crypto not isolated: latent factors price tech and profit","2023-24 crypto returns: equities lurk in latent space"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000288,"raw_usage":{"total_tokens":1484,"prompt_tokens":659,"completion_tokens":825,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":403,"completion_tokens_details":{"reasoning_tokens":741}},"tokens_in":403,"tokens_out":825,"duration_ms":8898,"temperature":1.0,"reasoning_tokens":741,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T11:02:09.927124+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-estimate the three-pass model on the same data with K=6 and K=8, and with alternative missing-data fill-in rules; if the Software, RMW, and Fear & Greed premia change sign or lose significance, the result is an artifact of the factor count. A sharper test: simulate returns from a known factor structure with missing observations and check whether the EM/PCA procedure recovers the true factors—if it does not, the mapped premia inherit that bias.","supporting_citations":[],"review_version":1}