At 0.152 RMSE, HAR-RV is the paper's second-best model and its real hurdle. The Corsi specification uses daily, weekly and monthly realized-volatility averages. The Transformer scores 0.179, CNN-LSTM reaches 0.165, and VMD-LSTM records 0.158, with variational mode decomposition preprocessing feeding an LSTM. Two published hybrids and a standard Transformer all failed to beat the linear benchmark against which the paper measures its architecture.

One disclosure belongs up front. We are building our own version on US equities and liquid ETFs, using price history from about 2010. We have no FX series and no 5-minute data extending back to 2000. The discussion below does not replicate the paper's numbers.

Agal's architecture

FinAnalyzeNet is presented as a hybrid deep learning framework for forecasting realized volatility. Its target is realized volatility calculated from 5-minute returns. The economic argument runs through VaR sizing, option pricing and volatility-target overlays. Hence the paper places a 4.9% VaR violation rate against a 5% nominal beside the RMSE table, along with a 1.42 Sharpe.

A learned gate combines three components. A temporal convolutional network extracts features across several time scales using six residual blocks, dilations 1 through 32, kernel size 3 and 64 filters. Three bidirectional LSTM layers, containing 256, 128 and 64 units, process those features forward and backward. Eight attention heads with key dimension 64 assign weights to time steps and features. The sigmoid-gated fusion layer in Eqs. 2 and 3 blends the three outputs. A mixture density head then produces point forecasts and intervals at 1, 5 and 20-day horizons.

The input set is broad. It includes returns over 1, 5, 10, 20 and 60 days, realized skew and kurtosis, the usual technical set, volume, Amihud illiquidity, macro series from FRED and OECD, VIX and put-call ratios, plus BERT-scored news and social sentiment. Recursive elimination selects the features using 2000-2017 only. They are then frozen, which is the right sequence, as the paper explicitly states.

The data run from January 2000 through December 2023 and cover 65 instruments: five equity indices, 50 large-cap stocks with a $10bn minimum, four FX pairs and six commodities. The total exceeds 350,000 daily records. Training covers 2000-2017, validation covers 2018-2020, and testing covers 2021-2023.

Across that test window, FinAnalyzeNet posts 0.134 RMSE against 0.152 for HAR-RV. The reported improvement is 11.8% (DM 3.42, p<0.001). Directional accuracy reaches 84.7%, compared with HAR-RV's 80.3%. At the 95% VaR level, the violation rate is 4.9% against a 5% nominal. The paper reports a Sharpe ratio of 1.42 without naming its window.

Where does 1.42 come from?

The Sharpe appears twice, in one clause of the abstract and one clause in the caption of Fig. 12. The evaluation section lists "trading strategy performance (volatility trading returns, hedging effectiveness)" as a category. We did not find the instrument or the mapping from forecast to signal. Nor did we find the rebalance schedule, gross or net exposure, a leverage or vol-target figure, turnover, transaction costs, or the Sharpe of a naive volatility strategy over the same 2021-2023 window.

A realized-volatility forecast needs a tradable implementation before it becomes a return stream. The vehicle could be variance swaps, delta-hedged straddles, VIX futures, or a volatility-target overlay on a directional book. Their costs and capacity differ. The paper identifies none of them.

We raised the same complaint about a genuinely good forecasting paper, the 164-year news index. Every monthly horizon improved there, while the trading claim depended on a spread too thin to survive costs.

Eighty point three percent deserves attention

A reported 84.7% directional accuracy would look extraordinary for return direction. This target is the direction of volatility, which is persistent. HAR-RV already achieves 80.3% on the same test set, leaving 4.4 percentage points attributable to the architecture on this metric. GARCH(1,1), at 58.3%, is the result that calls for an explanation.

The RMSE difference has a similar interpretive limit. Eq. 1 establishes daily standard deviation as the scale. The paper does not specify whether returns are expressed in decimals or percent, or whether annualization is applied. Without that information, the 0.018 difference cannot be translated into a position-sizing error or option mispricing. That translation is the point for risk work.

What the ablation actually credits

Removing the TCN costs 18.7% RMSE.

Removing the BiLSTM costs 12.3%, while removing attention costs 8.9%. Taking out gated fusion costs 4.2%. Agal says plainly that the latter result is not significant (DM 1.92, p = 0.08). The convolutional extractor therefore makes the larger single contribution, even though the recurrent component receives the headline emphasis.

The bidirectional pass also raises the causality question. The paper addresses it directly by masking the backward LSTM and attention to the window ending at t-1. It also trains a strictly unidirectional variant of identical capacity. That version records 0.138 RMSE versus 0.134 for the bidirectional model, giving bidirectionality a 3.2% gain (DM 2.18, p = 0.015).

I take the causal claim at face value. Its validation remains indirect, relying on no performance discontinuity at regime boundaries rather than a formal leakage test. About a fifth of the 0.018 RMSE advantage over HAR-RV comes from the feature that required the masking argument.

The advantage over HAR-RV changes sharply by regime. It is 8.9% when VIX is below 15, 16.3% above 25 and 19.1% during transitional periods. Bai-Perron breaks fitted on an expanding window provide ex-ante regime labels using only data through t-1. Those are precisely the periods in which a volatility-control overlay would need to earn its keep.

The bookkeeping problem

The paper's own dates show severe overlap among the walk-forward folds. Folds 4 and 5 test on identical dates, from 2023-01-03 to 2023-12-29. Fold 3's test window contains both. Fold 5's validation window, 2022-01-03 to 2022-12-30, is fold 2's test window verbatim.

The headline results include standard deviations calculated from those folds: 0.134 (sd 0.008) for FinAnalyzeNet against 0.152 (sd 0.009) for HAR-RV. The 0.018 gap is roughly two standard deviations wide, based on folds that share a test year. The reported dispersion therefore understates the real thing.

Table 2 gives multi-scale improvements of 11.8%, 12.4% and 15.6%. The surrounding text and conclusion instead report 10.1%, 13.2% and 15.8%. The abstract claims twelve benchmark models, while Table 1 lists eleven. Two different repository URLs appear. Raw data cannot be redistributed, leaving only a synthetic generator. The DM tests are one-sided.

At the 95% level, the Model Confidence Set retains every model statistically indistinguishable from the best. FinAnalyzeNet alone survives from the full comparison set, an unusually clean result.

The part we can test

We are building a version on US equities and liquid ETFs. Because our price history begins around 2010, the paper's 23-year span remains out of reach. We also have no FX spot or FX futures, so we cannot test the 10.2% currency-pair result. The claim of 5-minute bars back to 2000 for 50 individual names is beyond data we can match. Our features will instead use daily bars or 1-minute bars aggregated in code.

Our answer will therefore cover a narrower question. On a licensable US universe, can a six-block TCN with a bidirectional LSTM and eight attention heads beat HAR-RV by enough to pay for itself? The paper reports 18.7 hours of A100 training and 38 ms per inference. Its ablation table assigns the TCN the largest single contribution, with RMSE worsening by 18.7% when it is removed, compared with 12.3% for the recurrent core.

A net-of-cost Sharpe tied to a stated instrument and rebalance schedule over 2021-2023 would change my reading.