Bitcoin's reported forecast accuracy is too strange to take into a trading book without a backtest. Read literally, the row in Fu and Abdulabbas's results table describes either a daily return forecast explaining about 37% of test-period variance or a mislabelled target. The regularization idea is new. Evidence that it helps Bitcoin rests on one row in one table, and that row needs explaining.
Penalties give the heads their roles
The encoder is a standard transformer with 4 layers, 64-dimensional embeddings and feed-forward width 256. The authors use three attention heads, each with its own penalty. For the volatility head, squared Frobenius distance pulls the attention matrix toward a prior that weights past days by the proximity of their GARCH(1,1) conditional variance to the current day's. Those weights are softmaxed over the row. The trend head follows the same procedure, substituting a recursively computed one-sided trend for variance.
The mean-reversion head has a different penalty. It pulls the attended output toward a scaled trailing moving average. This directly constrains the output; attention weights are constrained only through their effect on it. Validation selects penalty weights of 0.3, 0.2 and 0.1.
The paper carefully describes its precautions against look-ahead. Only the training partition is used to fit GARCH parameters. Each prior stays within its own causal window, with normalization over past positions, and z-scoring uses training statistics. The datasets are front-month VIX futures returns from 2008 to 2023, the 128-series FRED-MD panel from 1959 to 2024, and Binance BTC-USD daily log-returns from 17 August 2017 to 31 December 2023. Economic-Loss Guided Multi-Head Self-Attention (ELG-MHSA) records the lowest NMSE on all three. NMSE divides test MSE by test-target variance, with results averaged over 1-, 3- and 6-step horizons.
For interpretability, the authors report a correlation of 0.72 between the volatility head and normalized GARCH variance. The trend head's correlation with the causal trend is 0.65. Both apply "in the displayed window". Training a head toward a target leaves it resembling that target, so the correlations show that the penalties took hold. Whether forecasts depend on those heads remains a separate question, raised by Serrano and Smith, whom the authors cite.
What does the Bitcoin gap establish?
On BTC-USD, ELG-MHSA's average NMSE is 0.63, compared with TFT's 0.70. Vanilla multi-head attention records 0.74, EGARCH 0.89 and GARCH(1,1) 0.91. ELG-MHSA is thus 10% below TFT and about 15% below vanilla attention. With 60-day input windows, the split comprises 1,233 training days, a 365-day validation year and a 729-day test set covering 2022 and 2023.
Every number comes from a single run using seed 42. The authors disclose this plainly and state that "no standard errors or statistical significance claims are attached". They describe the results as descriptive "unless accompanied by formal uncertainty estimates". None are supplied. Yet the conclusion says ELG-MHSA consistently outperforms TFT.
The conclusion qualifies the value of its convergence and regime results with "if confirmed in repeated-run studies". It places no equivalent condition on outperformance, or on the claim that the advantage grows monotonically with forecast horizon. Under the authors' own rule, the 0.07 gap remains a single draw.
The target is the larger problem. If it is the return, vanilla attention's 0.74 already implies that the model explains 26% of variance out of sample. That 0.74 averages the 1-, 3- and 6-step horizons. The series is described as daily log-returns, but we did not find a sentence specifying whether the forecast target is the return itself or its variance. GARCH and EGARCH model variance. They nevertheless receive scores on the same NMSE measure, and we could not find how their forecasts were mapped onto the target.
If returns are the target, the headline belongs to the plain transformer: it explains about a quarter of daily BTC return variance over 2022 and 2023, averaged across those three horizons. ELG-MHSA becomes a footnote. If the target is squared returns or volatility, those scores cease to imply return predictability. The trading application shifts to sizing from direction.
Table 2 includes neither a zero forecast nor a linear model, although the authors cite Zeng et al., whose title questions whether transformers are effective for time series forecasting, as evidence of transformer effectiveness.
Evidence from other markets
The ablation uses FRED-MD. At the 1-, 3- and 6-step horizons, the full model records 0.72, 0.79 and 0.85. Removing the economic losses raises those scores to 0.78, 0.88 and 0.97. A single loss applied to every head produces 0.75, 0.83 and 0.91; random pattern assignment produces 0.80, 0.90 and 1.02. The authors say random assignment underperforms the unregularized baseline "at longer horizons". Their table shows it doing so at all three.
The result is troubling. All three heads have the same architecture, so their roles come from their losses. In the paper's example of the variant, a volatility prior is "assigned to a head that receives a trend regularization loss". This describes a mismatch between prior and loss, beyond simply relabeling heads. The paper never specifies how that mismatch is implemented.
The authors attribute the gap to a volatility prior reaching a head "that should specialize in trend persistence". Either the variant alters training in an unspecified way, or the one-step 0.08 gap (0.72 against 0.80) falls within what seed variation could produce. A single seed cannot rule that out. It cannot distinguish the effect of the mapping from initialization luck.
Another inconsistency appears in the no-loss comparison. The paper describes that variant as effectively MHSA, yet its average is about 0.88, matching TFT's FRED-MD score. MHSA scores 0.95. We also did not find which of the 128 series serves as the forecast target.
The horizon result, 0.62 against TFT's 0.75 at 12 steps, comes only from VIX. So does the regime result of 0.45 against 0.55 when VIX exceeds 30. Bitcoin appears only in Table 2. The ablation therefore concerns FRED-MD, while the 12-step and regime results concern VIX. Fig. 3 leaves the source series for its alignment profiles unspecified.
A Bitcoin backtest must supply the economics
The paper reports forecast error alone, with no returns, Sharpe, turnover or costs. The relevant test would retain the same 60-day window and encoder, then compare the paper's 0.3, 0.2 and 0.1 weights with all three weights set to zero. It would repeat training across seeds and compare the resulting signals after costs. Beating that zero-weight twin would earn the auxiliary loss its 15% per-epoch compute overhead. Until then, 0.63 against 0.70 remains a forecasting curiosity whose target is unexplained.