A simulation seeded on the same data it later resembles offers no evidence about next week's volatility. Takaishi largely acknowledges the limitation.

We cannot test this paper's method. It contains no quantum hardware, and any strategy available to us would require a classical substitute. Testing that substitute would say nothing conclusive about quantum circuit learning.

The data are Bitcoin prices from Bitstamp, sampled every five-minute from 31 August 2020 to 1 March 2023, for 913 days. Daily realized volatility is the sum of the 288 intraday squared log returns. The previous day's return is rescaled to [-1/c, 1/c], while the previous day's RV is rescaled to [0, 1/c], with c = 1.2. Those values become rotation angles on a single qubit. A trainable unitary with three free parameters follows, and the probability of measuring |0> supplies tomorrow's RV. Qiskit computes the state vector on a classical machine. Takaishi claims no speedup.

Section 2 casts GARCH(1,1) as a truncated Taylor expansion of an unknown volatility map, with alpha, beta and omega corresponding to the expansion coefficients. Later variants, including QARCH, EGARCH and the rational GARCH family Takaishi himself has worked on, amount to different guesses about the functional form. The parameterized circuit avoids choosing one in advance.

Once fitted, the circuit generates 50,000 steps. Returns are constructed as R_t = RV_t^(1/2) times a standard Gaussian shock, then the resulting pair goes back into the model as its next input. Multifractal detrended fluctuation analysis is applied to the synthetic path. It uses rolling windows of 913 observations stepping 365, cubic detrending, q from -5 to 5, and scaling fits for s >= 20.

Every observation enters the fit

The paper states: "we do not explicitly divide the data into training, validation, and testing subsets." COBYLA minimizes the loss over all 913 days. Takaishi describes the goal as generative. Even so, the resulting 50,000-step path is an autoregressive loop whose parameters have seen every observation.

Several random COBYLA starts were attempted. Takaishi reports that "while the algorithm generally exhibited good convergence, we observed poor convergence in certain cases depending on the initial values." The displayed loss curve comes from one converged case, and we assume the synthetic series uses those same parameters. Similar loss values across the converged runs are the relevant defence. The reader still gets one trajectory from an unstated number of attempts.

Why do the returns look random?

Rolling h(2) for the simulated returns stays near 0.5. For the empirical input, h(2) = 0.504(5), compared with 0.498(11) after shuffling. The figures agree.

They have to.

Each simulated return is RV^(1/2) times an iid Gaussian. Signed returns are therefore serially uncorrelated by construction, regardless of the volatility multiplier. Takaishi says the result "simply confirms the generation of return time-series data based on Equation (34)." Section 6 thus disarms any attempt to treat the abstract's near-random return dynamics as evidence for the circuit.

The synthetic returns also appear multifractal, though the paper itself gives reason for caution. Takaishi generated 200 Gaussian series of length n = 913 and subjected them to the same MFDFA. Those series show q-dependent h(q) and a finite-width f(alpha): monofractal input produces multifractal output. Section 6 interprets the comparison favourably because the Gaussian series appear less multifractal, leaving possible multifractality beyond finite-size effects in the simulated series. Section 7.2 later assigns the multifractality that survives shuffling to those same finite-size effects. The excess is never measured through an f(alpha) width, a dispersion figure or a test. The abstract's claim depends on visually comparing two sets of identically scaled axes.

Roughness in the increments

The log-RV increment series carries the paper's genuinely non-trivial result. Empirical IRV has h(2) = 0.155(2), which rises to 0.516(11) after shuffling. Its anti-persistence comes from temporal ordering. Across the synthetic windows, IRV lies roughly between 0.05 to 0.10 and averages about 0.07. Shuffling again moves it toward 0.5.

A three-parameter bounded map is under no constructional constraint to create anti-persistent volatility increments. Takaishi makes that case in the conclusion, fairly.

The simulated roughness falls below the empirical 0.155, and we did not find a test of the difference. Proxy choice also matters sharply. On the same days, the empirical absolute-return increment series, IAR, records h(2) = 0.030(1), while IRV gives 0.155(2). Roughness changes by a factor of five for the same asset and period. Hurst estimates in these data depend heavily on the number of intraday samples used for the proxy. That agrees with earlier work Takaishi cites on sampling frequency and limits the weight a single roughness estimate can carry.

In the one plotted window, shuffling synthetic IRV pushes h(2) toward 0.5 while leaving the h(q) width unchanged. The q-dependence remains after temporal ordering is destroyed. Takaishi cites recent work arguing that genuine multifractality arises only from nonlinear temporal correlations, then declines to infer its origin from the shuffle. Fair.

The tail miss

Simulated return kurtosis reaches 15.7(33), against an empirical 5.79(60), a factor of 2.7.

Takaishi reports the miss and describes it as a possible failure to capture some properties of the data. Sections 7.1 and 7.4 offer the same answer: a single-qubit, three-parameter circuit has limited expressive power, while deeper multi-qubit circuits or data re-uploading may improve it. With no benchmark in either direction, there is no way to know whether extra depth would repair the tails or simply fit them. For VaR or option pricing, this is the paper's most consequential figure.

The discrepancy is confined to return levels. IRV and IAR kurtoses are close at 4.43(87) versus 4.55(7), and 4.40(42) versus 4.35(10). The return-level mismatch emerges where the RV^(1/2) times Gaussian mapping operates.

What would change our view?

The absent benchmark decides the paper's usefulness, and Takaishi recognizes the omission. Comparison with GARCH, HAR and Realized GARCH is "beyond the scope" and left for future work. We did not find any forecast error metric: no QLIKE and no MSE on a holdout. Readers receive a loss-versus-iteration chart and a visual overlay of fitted RV against input RV. The overlay captures the broad shape.

Run rolling out-of-sample one-day-ahead RV forecasts across the same 913 days, comparing the circuit with HAR-RV under QLIKE and MSE. HAR uses three coefficients on daily, weekly and monthly RV averages.

The evidence covers a single asset on a single exchange over 2.5 years of one crypto regime. Missing five-minute intervals are filled by carrying the previous price forward. We would expect that treatment to suppress measured variance on thin days, and it could affect increment scaling. A favourable QLIKE result against HAR on a genuine holdout would change our view.