Weekend crypto variance improved Monday variance forecasts by 11 to 22 percent after April 2024 for the S&P 500, gold and a Treasury bond fund. That belongs in a risk model; nobody should book it as alpha. Ebadollahyvahed and Umeorah report QLIKE losses declining from 0.695 to 0.545 for the S&P 500, from 3.182 to 2.845 for gold and from 0.424 to 0.350 for IEF. They score forecasts. The paper reports no strategy returns, Sharpe ratio or costs.
Making money from a better variance forecast requires sizing. That sizing earns its keep only when the Mondays it cuts are ones you would rather have missed. The targets also differ: a cash index, a continuous COMEX futures series and a 7-10 year Treasury ETF. Gold futures trade different hours from the other two.
What arrives before Monday's bell?
The authors start with daily OHLC bars for Bitcoin, Ether and XRP, computing a Garman-Klass variance for each. They add Saturday and Sunday, take logs, then give the three coins equal weight. They call this common weekend crypto variance factor W. Applying the same calculation to Friday produces F.
Each traditional asset has a Monday-only HAR benchmark, a regression using its own log variance from the previous session and averages over 5- and 22-session windows. The authors then add W and F.
The identification depends on timing. Daily crypto bars end at 00:00 UTC, putting the weekend window between Saturday 00:00 and Monday 00:00 UTC. US cash trading finishes on Friday at 20:00 or 21:00 UTC, then resumes Monday at 13:30 or 14:30 UTC. None of the information in W could therefore be traded in the US cash equity session before Monday's open. Friday overlaps an open session, making F a placebo. If W merely captures recent crypto turbulence, F should forecast Monday equally well; their correlation is 0.67.
The authors identify two limits to that timing argument. COMEX gold resumes trading on Sunday evening, allowing traders to act during the Sunday bar's final hours. And the Monday target measures the within-session range, leaving out the Friday-close-to-Monday-open gap. This suits an overlay entering at Monday's open. It leaves out the gap borne by anyone holding across the weekend.
After January 2024, all nine crypto-asset pairs are significant
The mean weekend coefficient across the nine coin-asset pairs starts at -0.03 before COVID and reaches 0.03 in 2020-21. Neither period has a positive, significant pair. In 2022-23 the mean reaches 0.14, with 2 of 9 pairs significant. After January 2024 it is 0.35, and all 9 are significant. A panel estimated over rolling two-year windows settles near 0.30 in 2025-26.
For post-2024 regressions using the weekend factor alone, W's coefficients are 0.256 for the S&P 500, 0.447 for gold and 0.246 for IEF. Each has p at or below 0.005. Using Friday alone gives coefficients of -0.015, -0.017 and 0.034, with p-values ranging from 0.521 to 0.883.
Under the authors' scaling, a one-standard-deviation change in W (0.97) accompanies fitted Monday variance increases of 28, 54 and 27 percent. Incremental R² comes to 3.9, 6.2 and 7.0 points. Adding each coin's residual produces no significance at 5 percent, placing the information in the common factor. The date-clustered panel coefficient is 0.38, with a t of 5.6 across 116 Mondays.
A Diebold-Yilmaz connectedness estimate on the same six assets after 2024 assigns individual coins only 1 to 7 percent of traditional assets' forecast-error variance. This decomposition measures how much forecast error comes from the other assets. Its rolling 250-day index shows no break around the ETF approvals. The authors' explanation is plausible: pooling weekdays dilutes an effect confined to Mondays.
A short window bears the weight
There are 115 or 116 Mondays in the post-2024 regressions and 103 or 104 in the out-of-sample test. The authors acknowledge the constraint, saying their emergence result depends on a limited number of observations.
Yet the abstract calls the predictive content "significant thereafter". In-sample p-values look decisive; the forecast tests supporting that out-of-sample claim are less convincing. The one-sided Diebold-Mariano p-values are 0.066 for the S&P 500, 0.050 for gold and 0.042 for IEF. The authors acknowledge that statistical evidence "varies across markets". On a risk desk, I would give the DM test more weight. Newey-West with 5 lags and roughly 115 weekly observations leaves little to work with. As the authors also acknowledge, clustering by date leaves serial dependence across Mondays untreated.
Pinning down the break is harder still. The sup-Wald procedure selects October 2023, with sup-W around 8.8, significant at 10 percent and borderline at 5. Changing the range estimator shifts the date to December 2022 with Parkinson (6.8), or April 2024 with Rogers-Satchell (10.1). The appendix itself cautions against taking October 2023 too literally.
The appendix distinguishes the level from its timing. Every estimator retains the post-2024 level, while the authors themselves describe the timing as less dependable. Parkinson yields 9 of 9 significant pairs; Rogers-Satchell yields 8 of 9. Across all variants, QLIKE declines by 9 to 25 percent. Those level estimates still come from the same short window.
Gold needs its own warning. The 54 percent effect is the largest, but before cleaning it has 131 zero-variance days under Garman-Klass and 367 under Rogers-Satchell. Gold's Rogers-Satchell log variance has a correlation of only 0.75 with Garman-Klass. Switching to Rogers-Satchell reduces its effect to 37 percent.
Do holidays or ETF flows explain it?
The relationship weakens in the 11 weeks when a Monday holiday delays reopening until Tuesday. Interaction coefficients are -0.45 for the S&P 500 (p = 0.01), -0.50 for IEF (p = 0.01) and -0.23 for gold (p = 0.36). An explanation based purely on accumulating information would predict at least as much signal after a longer closure. Here, there is less.
The institutional explanation also struggles. Interactions between W and standardised absolute spot ETF flows have p = 0.30 for the S&P 500 and 0.85 for gold. IEF's t is significantly negative at -2.6. The authors describe the flow proxy as "coarse and endogenous" and ask readers to treat the null as absence of evidence.
Neither finding invalidates the forecast.
A forecaster can use W if it works out of sample without knowing the mechanism. Persistence is harder to judge without that explanation. A coefficient of 0.35 over 2024-26 could describe a regime or a lucky draw; the missing mechanism leaves us unable to distinguish them.
Splitting the S&P 500 by VIX offers a clue for sizing. W's coefficient is 0.090 below the Friday VIX median and 0.675 above it (p = 0.0055). The authors caution that these p-values have no adjustment for six comparisons.
Our SPY, GLD and TLT adaptation
We cannot hold the index or GC=F, so we adapted the approach to SPY, GLD and TLT. Our TLT substitution brings long-duration exposure in place of the paper's 7-10 year IEF.
The design uses two matched ledgers. HAR determines sizing in one; HAR plus the weekend factor determines it in the other. Both fit an expanding window requiring at least 78 Mondays. We standardise the weekend factor using preceding Mondays alone and make the decision before 09:25 New York time. To the session forecast, we add overnight variance estimated from the last 60 close-to-open intervals.
Each ETF sleeve targets 10 percent annualised volatility, subject to a cap at its one-third baseline allocation. Reductions trade in Monday's opening auction. The next session's open restores baseline positions. We charge commissions of $0.004 a share, with a $1 minimum.
From 2020-01-01 to 2024-07-01, our book earned 16.53 percent in total, or a 3.47 percent CAGR. Its Sharpe was 0.35 and maximum drawdown was -24.38 percent. We report the aggregate book alone. Without a matched HAR-only result, we cannot measure the overlay's contribution. The paper supplies no Sharpe or return for comparison. It reports an 11 to 22 percent reduction in QLIKE from April 2024 to June 2026. These figures measure different things: our result is the whole equal-weight book's P&L, and theirs measures forecast loss.
Our Sharpe cannot tell us whether W improved the forecasts either. Between Monday-open trades, positions remain at baseline weights. Annualised book volatility was 10.66 percent, near the 10 percent target, and beta to SPY was 0.29. The drawdown therefore likely reflects TLT during the 2022 rate shock more than the overlay.
The dates impose another limit on what this run can say about the paper. Our sample overlaps the authors' out-of-sample period by about three months, roughly a dozen Mondays. Most of our Mondays belong to years in which they report 0 of 9 and 2 of 9 significant pairs. Our expanding fit starts in 2020, so pre-break observations likely dominated the weekend loading estimate, diluting it well below their 0.25 to 0.45.
Our pre-open data-availability requirement offers another possible explanation: it may have disabled the overlay on many Mondays. Both ledgers would then follow the same baseline schedule. We have yet to count eligible Mondays to test this possibility. Our one automated pass neither reproduces nor contradicts the authors.
A fair test would size positions on Mondays in a strictly walk-forward run beginning after the break. It would compare paired forecast losses using ETF prices and hold out 2026. If SPY's weekend loading in that window returns toward the near-zero values found in the paper for 2020-21 (a nine-pair mean of 0.03), the channel was a two-year episode. I would then stop using it as a risk input.
Our backtest stops at 2024-07-01, and everything after that date is deliberately left untouched so the same strategy can be checked out of sample later.