Bitcoin spot and futures volatility move nearly as one. The paper puts their connectedness close to the limit of its two-market model, though it gives traders little basis for acting on the reported spillover direction.
Kassamany, Naïmy and Jabbour study volatility passing between a cryptocurrency's spot market and its own futures market. A BTC future carries the same exposure as spot plus a financing leg. Cash-and-carry arbitrage should therefore tie the prices together and pull their volatilities into line. The authors estimate an ARFIMA mean and a GARCH-family variance for four daily return series: BTC spot, BTC futures, ETH spot and ETH futures. For each coin, a TVP-VAR then takes the two conditional-volatility series and updates its coefficients daily through a Kalman filter. Forecast-error variance decomposition supplies the spillover measures; dynamic conditional correlation (DCC) measures co-movement in those same volatility series.
The futures are standard contracts from LSEG Workspace and Datastream, with no perpetuals. Spot prices are volume-weighted aggregates from CoinMarketCap and CoinGecko. The BTC sample runs from December 2017 through 2022, with 1,276 daily observations. ETH runs from February 2021 to end-2022, with 269 observations. The total connectedness index (TCI) averages the share of each market's forecast variance explained by the other. It reaches 43.48% for BTC and 28.98% for ETH; the reported DCC coefficients are 0.92563 and 0.64726. Futures are the net transmitter for both coins, at +0.15 for BTC and +8.29 for ETH.
How much does Bitcoin's 43.48% tell us?
Each market's own shocks account for roughly half or more of its forecast variance in this bivariate model. The TCI consequently has an effective ceiling near 50%. BTC's 43.48% leaves spot and futures volatility looking almost like a single series, just as arbitrage would suggest.
ETH is harder to account for. Its two legs are tied by the same kind of arbitrage, yet its DCC is 0.64726 and its TCI is 28.98%. The abstract offers market maturity, liquidity and sample period as explanations. The DCC discussion adds Bitcoin's dominance, put by the paper at roughly 54% versus 18% for ETH. I would check the data construction before settling on that economic reading.
Precise filter settings, uncertain market clocks
Following Antonakakis, Chatziantoniou and Gabauer (2020), the authors specify the connectedness stage closely: lag 1, a 16-step generalized decomposition, a forgetting factor of 0.99 and a decay factor of 0.96. Their rolling-VAR comparisons use 100-, 150- and 200-day windows. AIC selects the GARCH models. For ETH, the authors use normal errors because Student-t and GED parameters proved unstable on 269 points.
Four gaps in the data construction matter for anyone trying to reproduce the result:
- Contract and roll. We did not find the exchange whose contract underlies the futures series or the method used to roll it.
- Timestamp. Spot is "aligned to the same daily frequency", but we did not find its cutoff time. Closing a 24/7 aggregate at midnight UTC and comparing it with an exchange settlement hours earlier can mechanically reduce same-day linkage and produce a spurious lead-lag. Both coins draw on the same spot and futures sources, so such a mismatch could lower measured linkage for each.
- Sample counts. BTC's 1,276 observations amount to about five years of exchange trading days. The ETH window spans roughly 470 exchange trading days (about 680 calendar days), while the sample contains 269 observations. We could not reconcile the count.
- Volume. One passage describes futures and spot volume as influencers and controls. Another says no controls enter any equation. We would follow the latter statement, since the authors describe their equations as the complete specification.
The fitted models raise separate concerns. In the ETH futures ARMA(2,2), the absolute AR(2) and MA(2) coefficients are 0.9902 and 1.0053. That leaves the model near a unit root, with its MA component in non-invertible territory. I would refit at a lower order. For BTC, the authors acknowledge residual autocorrelation: the Ljung-Box p values are 0.0193 for spot and 0.0014 for futures. They describe it as mild and first-lag and say it "does not affect the adequacy of the conditional-variance specification". That defence is reasonable for the variance stage, since those volatility series enter the TVP-VAR. Residual mean autocorrelation still reaches the squared residuals that drive them.
Follow the spillover table
Rows in the paper's spillover table receive shocks. In the BTC spot row, 43.56 comes from futures. The accompanying text instead says 43.56% spills from spot to futures. ETH has the same reversal: its spot row receives 33.13 from futures, while its futures row receives 24.83 from spot. The text assigns 33.13% of futures variability to spot.
Under the prose's reading, ETH spot would transmit on net. The paper's NET row gives futures +8.29, consistent with the standard table convention. I would trust the table.
For BTC, a net of 0.15 beside cross shares of about 43 gives a trader little directional signal. The network plot nevertheless calls the BTC_F to BTC_S edge a "significant spillover effect of futures trading on spot market volatility". ETH's +8.29 is the sole NET figure with much directional weight, and it comes from 269 observations whose timing remains unclear.
After the futures launch
The results section concludes that "the trading of BTC futures is associated with an increase in spot market volatility". Its spillovers, however, are estimated only over periods when both markets exist. We found no fitted model for the collected December 2016 to December 2017 BTC pre-period or for the ETH pre-period from 1 February 2020 to 28 February 2021. Augustin, Rubtsov and Shin, cited by the paper, treat the same December 2017 launch as an event. This paper attempts no comparable event study.
The abstract also says strong co-movement "supports the use of futures for hedging spot exposure". We found no hedge ratio, hedging-effectiveness figure or cost estimate. The authors acknowledge that the 0.92563 DCC has no standard errors and serves as a descriptive measure only.
Even the BTC-ETH contrast strains the design. Sample period appears among the abstract's explanations for the gap, and the authors say ETH's TCI is "not directly comparable to Bitcoin's longer, multi-cycle sample". Yet the results later attribute the DCC gap to Bitcoin's dominance. Comparing 2017-2022 with 2021-2022 cannot disentangle dominance from regime.
We could not test the result ourselves. We have no cryptocurrency futures series; spot prices and crypto ETFs cannot replace the futures leg. Our crypto history starts around 2018, after the paper's BTC pre-period begins in December 2016.
The useful result is a reading to monitor. BTC's full-sample TCI of 43.48% lies near the 50% ceiling, making a sharp fall from that level worth watching. The authors likewise suggest that regulators monitor changing interconnectedness between futures and spot markets. A timestamp-matched re-estimate that still puts ETH's DCC near 0.65 would change my view of its weaker linkage.