Research
NEST: Stale-Robust Tokenized-Stock Measurement for a Crypto-Treasury LETF Stack
Separates real predictive signal in tokenized-equity marks from mechanical shared-history fit, supplying a bound that kills spurious off-hours measurement claims.

- arXiv
- 2610.21491
- Published
- 2026-10-15
- Authors
- Doohwi Cha, Minsuk Sung, Minjae Lee, Hyeonjun Yeo, Hanhee Kim, Seunghan Son, Donghwa Seo, Juyeong Lee
AI summary
Overview
Research area: Trading and market microstructure, specifically tokenized equities (tracker certificates on U.S. stocks that trade on crypto venues) and the Bitcoin → crypto-treasury stock → leveraged ETF exposure chain.
Technical level: Advanced. The paper is built from population regression identities, a variance-clock bound, partial correlation decompositions, a noncentral F finite-sample frontier, a Newey–West HAC test, and a daily-reset leveraged-ETF expansion. The framing is accessible, but the machinery is econometric.
Scope (one sentence): The paper builds and tests NEST, a protocol that decides whether an off-hours tokenized-stock mark carries genuine conditional information about the next U.S. opening gap, or whether its apparent fit is an artifact of shared price history and stale timing.
What This Paper Is About
Tokenized equities such as MSTRx trade overnight, on weekends, and on exchange holidays, when the underlying listed share does not trade — so their prices could in principle measure exposure while the cash market is shut. The problem is that a token return and the next opening gap both start near the previous official close, so they share the same price move, and a regression of one on the other produces a high fit even if the token says nothing about what happens after its own timestamp. The paper's goal is to separate that mechanical shared-history fit from the token's genuine incremental value beyond public information observable at a 09:25 ET cutoff, and then to decide whether such a mark is admissible for use in a Bitcoin–MSTR–leveraged-ETF exposure stack.
Key Contributions
-
Conceptual — the mechanical-fit diagnosis. A shared-history identity (Proposition 1) shows that the population slope and R² of the opening gap on the token return are governed by the shared previous-close-to-mark move relative to the post-mark move and mark noise, while the covariance between the token and everything after its mark is exactly zero. A variance-clock bound (Proposition 2) caps R² at V(θ)/V(1), the share of closed-window variance realized by the time of the mark — a late mark can fit almost perfectly with no information about the remainder.
-
Measurement — PDI and an exact finite-sample frontier. The partial determination index (PDI), the squared partial correlation of gap and token given public controls, is shown to equal (R₁² − R₀²)/(1 − R₀²) (Lemma 1). Theorem 3.1 gives its exact finite-sample distribution as a noncentral F, yielding a critical value c_α(n,p) in PDI units; Corollary 1 supplies the power function and a minimum detectable PDI, θ_π. Lemma 2 shows that a positive PDI against a noisy benchmark means improvement over that benchmark, not exclusive information origin.
-
Operational — admissibility and stack transmission. A four-step admission protocol (clock, conditioning, decision, admissibility) combines the PDI frontier with a HAC check and audits of product rights, price provenance, timestamps, a named venue, and execution costs. The stack layer adds a daily-reset identity (Lemma 3) with its variance-drag term, an effective MSTR-on-Bitcoin slope κ = λ_ℰ + ρ_m with λ_ℰ = qC/(qC − D), and a one-reset wipeout threshold (ΔC/C)_wipe = −1/(βλ_ℰ), compared to −1/(βκ) when the premium moves with Bitcoin.
-
Empirical — a same-clock design. A clock that structurally cannot see the open (all inputs observable by the 09:25 ET cutoff, with a stricter variant requiring both token marks to precede their targets), applied to MSTRx plus 694,246 reconstructed Gate MSTRX–USDT trades and a cross-name check on NVDAX, TSLAX, and AAPLX.
Main Findings
-
Raw fit is almost entirely shared history. On 222 strict-cutoff windows from July 2025 to June 2026, the token-only regression has R² = 0.916. Once Bitcoin, mark age, and MSTR's pre-09:00 premarket proxy are controlled, the token's absolute gain is ΔR² = 0.014, taking the full model to 92.3% from 90.8%.
-
The conditional increment survives, but is modest. MSTRx explains 15.7% of the gap variation left by Bitcoin, mark age, and the premarket proxy (PDI = 0.157, F = 39.6, HAC t = 4.61, 1% frontier 0.031) on the 218 common windows. Against Bitcoin and mark age alone, the increment is ΔR² = 0.133 with PDI = 0.620 (HAC t = 10.92, frontier 0.030).
-
The token is complementary, not dominant. Token and proxy correlate at 0.970, and the reverse increment of the proxy given the token is PDI = 0.082 — the two prices overlap substantially without either spanning the other.
-
The clock audit holds. Requiring every token mark to precede 09:00 gives PDI = 0.140 (N = 214, HAC t = 4.36); both marks by their targets gives 0.170 (N = 215); restricting to Bitcoin end marks at or before 09:25 gives 0.168 (N = 106); marks within 30 minutes gives 0.168; winsorizing at 1%/99% gives 0.165; a 09:00-bar-open proxy gives 0.150. The composite end marks fall between 08:56 and 09:15 ET, and 214 of 218 precede 09:00 by 0.85 to 3.15 minutes (median 0.97).
-
Trade-level reconstruction on Gate is stronger than the composite. Using the last trade by 08:59 ET, N = 216 gives ΔR² = 0.023, PDI = 0.252, HAC t = 6.85, frontier 0.031. Without the premarket control, PDI = 0.792 (N = 225, HAC t = 26.78).
-
Results are product specific across xStocks. NVDAX clears (N = 204, PDI = 0.169, frontier 0.033, HAC t = 4.85), AAPLX clears (N = 224, PDI = 0.092, frontier 0.030, HAC t = 5.16), and TSLAX does not (N = 197, PDI = 0.010, frontier 0.034, HAC t = 0.80).
-
Weekends carry the same increment, measured with less precision. Weekday windows give PDI = 0.122 (N = 170) and weekend-or-holiday windows 0.112 (N = 48); the token-by-weekend interaction has HAC t = −0.14 (p = 0.89), with token loadings of 0.52 on weekdays and 0.49 on weekends. The weekend estimate falls below its own frontier of 0.145 only because 48 windows detect a PDI of only about 0.21 at 80% power and 1% level, and at the weekday value the test has 46% power.
-
Sample and clock sensitivities do not overturn the broader result. Lag limits of 30 to 60 minutes, exclusion of early-close starts, and winsorization leave the Bitcoin-and-age PDI between 0.613 and 0.629. Requiring all four endpoints to precede their targets cuts the sample to 67 windows and gives 0.445 against a frontier of 0.101. Splitting at the January 26, 2026 vintage of CoinGecko's aggregation methodology gives 0.634 before and 0.559 after. Without the premarket proxy, the increment is 0.784 for multi-day closures and 0.824 for the 41 Friday-to-Monday windows.
-
Power matters for interpretation. At the 1% level and 80% power, 218 windows detect any PDI above 0.052, whereas 48 windows detect only PDIs above 0.21.
-
Admissibility is separate from statistical significance. MSTRx is a variable-denomination tracker certificate issued by Backed Assets (JE) Ltd. with no voting rights, segregated MSTR collateral, and primary creation/redemption subject to KYC, whitelisting, fees, and minimum sizes; every observed MSTRx confidence value equals 0.99, so that field screens availability rather than liquidity. The executable band of Proposition 3 is therefore described as a friction benchmark, not an enforceable arbitrage, for most holders.
Methodology in Plain English
The authors start by writing the opening gap as two pieces — the move from the previous close through the token's mark, plus the later move from the mark to the open — and the token return as that same first piece plus measurement noise. Under this decomposition they can prove algebraically what the raw regression must look like, and show that the token has zero covariance with the post-mark piece. They then replace calendar time with a "variance clock" that measures how much of the closed-window's volatility has already occurred by the mark, which bounds how high raw R² can get from timing alone.
To ask what the token adds, they take the gap and the token return, strip out everything public that is observable by the 09:25 cutoff (an intercept, the Bitcoin return, the age of the mark, and MSTR's own 08:00-ET premarket bar close), and compute the squared correlation of the two leftovers — the PDI. Because a sample PDI is noisy, they derive its exact distribution under the null and convert it into a critical value on the PDI scale itself, so a reader compares the estimated share directly with the share required for significance, plus a power calculation that says how large a PDI a given number of windows can detect. Every frontier decision is paired with a Newey–West HAC t-statistic with five lags.
The data design is deliberately backward-looking: token and Bitcoin marks are requested at the previous 16:00 ET close and at a 09:25 ET cutoff, five minutes before the open, and a stricter variant requires both marks to precede their targets. This replaces an earlier nearest-mark rule, which on the 232 availability-screened windows had selected 10 end marks after 09:25 — 8 of them at or after the 09:30 open. Prices come from DefiLlama's archive of CoinGecko's multi-venue MSTRx reference price, Financial Modeling Prep's official MSTR close and next open, and Yahoo Finance's hourly extended-hours chart for the premarket proxy; venue evidence comes from 694,246 MSTRX–USDT trades reconstructed from Gate's twelve official monthly trade files for July 2025–June 2026, with 24 order-book snapshots at the window endpoints for a cost check. Statistical clearance is then passed through a separate admissibility layer covering product rights, provenance, timestamps, and execution cost before the mark is used to update stack exposure diagnostics.
Why This Matters
Impact on research. The paper isolates a failure mode that plausibly affects a wide class of off-hours and asynchronous-market studies: when a measurement and its target share a common starting point, raw R² is uninformative, and the size of the artifact can be bounded in advance. It gives the tokenized-asset literature a testable conditional object (the PDI with its finite-sample frontier) rather than another raw-fit comparison, and it makes the noisy-benchmark interpretation explicit — a positive PDI means the mark beats the observed public benchmark, not that it originates the information.
Real-world applications.
- Monitoring Bitcoin-linked equity exposure overnight, on weekends, and on exchange holidays, with a defensible statement of how much the off-hours mark actually adds beyond the listed stock's own premarket price.
- Risk diagnostics for daily-reset leveraged ETFs whose underlier is itself a balance-sheet-levered Bitcoin claim, including the one-reset wipeout threshold (ΔC/C)_wipe = −1/(βλ_ℰ).
- Vendor and product evaluation: deciding whether a specific tracker certificate (MSTRx, NVDAX, AAPLX, TSLAX) is admissible for a given use, given provenance, timestamps, and execution costs.
- Study design for small samples, using Corollary 1 to check whether a subsample could have detected the effect at all before reading a failure to clear as evidence of no information.
Industry relevance. The results land in a specific place: the token is a complementary measurement, not a dominant price-discovery venue, and product-level differences (NVDAX and AAPLX clear their frontiers, TSLAX does not) mean admission decisions cannot be made at the asset-class level. The admissibility audits — tokenized tracker versus common stock, undisclosed composite weights, restricted creation and redemption — speak directly to desks that might otherwise treat a token print as an executable or enforceable price.
Future Directions
-
Out-of-sample durability. Section 5 describes expanding-window and frozen July 2026 evaluations testing whether the Bitcoin-and-age increment persists out of sample; the reported results here stop before those are detailed, so the stability of the increment over later windows remains an open question.
-
Which κ to use for the stack. The wipeout threshold depends on whether book leverage, the listed-session beta, or the slope implied by an admitted off-hours mark is used to update κ; Section 6.7 is referenced for this but its content is not in the material summarized here.
-
Making the executable band enforceable. Proposition 3's band requires locking, eligibility, and inventory that KYC, whitelisting, fees, and minimum sizes currently prevent for most holders; testing the band under actual creation and redemption access is left open.
-
Power in short subsamples. The weekend analysis shows how easily a true increment is missed when windows are few (48 windows, θ₈₀ ≈ 0.21, 46% power at the weekday value); extending the sample or pooling products to recover power is a natural next step.
Target Audience
Quantitative researchers and practitioners in market microstructure and crypto-asset research; desks and risk teams monitoring Bitcoin-linked equity and leveraged-ETF exposure outside U.S. hours; product and compliance reviewers evaluating tokenized-equity tracker certificates; and statisticians interested in partial-correlation inference and design calculations for small, asynchronously sampled datasets. Readers without econometrics background will follow the framing and the headline numbers but will need the propositions and the F-distribution results to use the protocol.
Note: the paper content supplied is truncated partway through Table 5, so later sensitivity results and Sections 7–9 (implications, limitations, conclusions) are not reported in this summary.
Authors’ abstract
Tokenized equities trade while listed markets are closed, so their marks could improve off-hours measurement along layered exposures that link Bitcoin, crypto-treasury stocks, and leveraged ETFs. A token return and the next opening gap, however, share the price move from the previous close to the token mark, so part of their apparent fit is mechanical. We develop NEST, a stale-robust protocol that separates this shared-history fit from conditional measurement value. A…
The full abstract on arXiv