MSTRx deserves a place on the pre-open risk screen. Buying $10,000 of it on Gate costs a median 72.0 bp, though, and the thinnest book had only about $436 on one side within 100 bp of mid.
The question before 09:30
MSTRx is a tracker certificate on MicroStrategy stock. Backed Assets issues it, and crypto venues trade it overnight, on weekends and on holidays. Cha, Sung, Lee and coauthors ask whether its off-hours mark explains MSTR's next 09:30 opening gap beyond the information available to a desk at 09:25: Bitcoin's move, the age of the token mark and MSTR's own premarket print. Bitcoin affects MSTR through its balance sheet; MSTR, in turn, affects a daily-reset leveraged ETF. A better overnight estimate of MSTR's Bitcoin sensitivity changes the Bitcoin drawdown that would wipe out a 2x fund in one reset.
The authors call their protocol NEST (Nested Exposure Slope and Thresholds). It measures the token's contribution after accounting for stale marks and shared price history. The central measure is the partial determination index (PDI): residualize the token return and the opening gap on the controls, then square the correlation between those residuals. That yields the share of remaining gap variance captured by the token. Each PDI faces a frontier derived from the partial-F test, exact under Gaussian errors, alongside a Newey-West t at five lags. The authors also check the token's legal status, the source and timing of its prices, and its trading costs. Their data cover July 2, 2025 to June 5, 2026, with 234 candidate windows and 222 in the strict sample. Token marks come from DefiLlama's archive of CoinGecko's multi-venue composite. A separate venue check draws on 694,246 Gate MSTRX-USDT trades.
Why is the raw fit so high?
Regressing the gap on the token return alone gives R2 = 0.916. Both returns start at or near the previous MSTR close. The composite updates roughly hourly, leaving its end mark usually at 08:59, a median 26.0 minutes before the cutoff. Consequently, the move from 16:00 to 08:59 appears in both series. The authors prove that the shared portion can generate fit as high as the share of closed-window variance realized by mark time, even if the token has zero covariance with subsequent moves.
A clock check found an actual leak. A naive nearest-mark rule selected a token mark after 09:25 in 10 of 232 windows; 8 of those marks came at or after the open. The primary design instead requires the end mark by the cutoff. Every token return reported below uses that repaired clock.
The premarket print takes most of it
Across 218 matched windows, Bitcoin and mark age explain 77.9% of the gap. Add the premarket proxy and the figure rises to 90.8%. MSTRx contributes another 0.014 of R2, bringing the total to 92.3%. Relative to the unexplained portion, its PDI is 0.157 (F = 39.6, HAC t = 4.61), above a 1% frontier of 0.031. The token and proxy correlate at 0.970, yet the proxy adds a PDI of 0.082 after accounting for the token.
Each still contributes information beyond the other.
Restricting the sample to the 214 windows whose token marks all precede 09:00 leaves a PDI of 0.140 (t = 4.36). Across clock variants it ranges from 0.140 to 0.170. Weekday windows produce 0.122; weekend-or-holiday windows produce 0.112, with an interaction p of 0.89. The weekend estimate falls short of its own 0.145 frontier. With 48 windows, that sample has 80% power at 1% only for a true share of at least 0.21. The result varies by product: NVDAX clears at 0.169 and AAPLX at 0.092, while TSLAX fails at 0.010 (t = 0.80).
The benchmark remains the weak point. Its premarket proxy is the close of a Yahoo 08:00-ET hourly bar, and every matched bar reports zero volume. The authors acknowledge the problem: if that proxy is noisy, a sharper one would reduce the PDI. They propose timestamp-matched consolidated quotes as the confirmatory test. The token mark at about 08:59 and the bar end at 09:00 are nearly simultaneous, which makes the quality of that comparison matter. I would regard the increment as settled if a consolidated-quote benchmark at 09:00 still yielded an estimated PDI above the 0.031 frontier in a sample powered to detect a true share of 0.052.
Gate's prints and its order book
Trade-level data from a named venue give the signal its strongest check. Gate's last trade by 08:59 clears the premarket-conditioned frontier with a PDI of 0.252 (t = 6.85, N = 216), versus 0.157 for the composite. At common timestamps, Gate adds 0.100 given the composite, which adds 0.057 given Gate. They largely convey the same information.
Trading the token at size is another proposition. Across 24 endpoint snapshots, Gate's median quoted spread is 19.1 bp. A $1,000 market order costs 23.7 bp to buy and 20.8 bp to sell; a $10,000 order costs 72.0 and 49.3 bp. In the thinnest snapshot, one side of the book contained about $436 within 100 bp of mid. Creation and redemption require KYC and whitelisting, and carry fees and minimum sizes, leaving most holders without an arbitrage band to rely on. As the authors put it, "Statistical informativeness is not executable value." I did not find a test of trading the information in MSTR premarket or in the LETF, where a desk would actually use it.
What did the holdout control for?
The forecasting results are large. Expanding windows reduce MSE by 59.8%, 57.6% and 61.3% relative to Bitcoin and age (Clark-West t of 5.46, 4.79 and 4.43). With coefficients frozen on June 5, a July 2026 holdout gets 21 of 22 gap signs right and cuts MSE by 77.8%. These tests use the Bitcoin-and-age specification. Its PDI is 0.620, falling to 0.157 when the premarket print enters. The authors flag the gap themselves and leave a premarket-conditioned holdout for a later sample.
They also report that the frozen model's RMSE of 0.0084 cannot be distinguished from the raw token's 0.0081. The Bitcoin-and-age model has a July RMSE of 0.0178. Both token-based forecasts roughly halve that error; the raw token does so on its own. PDI measures incremental fit through the mark. NEST, the authors say, does not forecast post-mark shocks, and decay of the conditioned 0.157 remains untested.
We could not rerun any of this. Our data contain minute bars for MSTR, US leveraged ETFs and spot Bitcoin, but lack timestamped MSTRx marks and Gate trade prints. There is therefore no token return for us to residualize.
The exposure update is the use that holds up. The token's closed-window Bitcoin slope is 1.077 (s.e. 0.043), near book leverage of 1.19 and well below the 60-day listed beta of 1.474. For a one-reset 2x wipeout, the corresponding Bitcoin falls are 46.4%, 42.0% and 33.9%. The token slope lies at the gentle end of that range. It gives a risk desk one estimate to weigh against book leverage's -42.0%; the spread across those links is worth seeing before 09:30.