FinAnalyzeNet's 1.42 Sharpe arrives with no trading rule attached

HAR-RV sets the standard that matters in Agal's paper. The regression, which uses daily, weekly and monthly averages of realized volatility, ranks second across the entire field with test-period RMSE of 0.152. It beats the Transformer at 0.179, CNN-LSTM at 0.165 and VMD-LSTM at 0.158. Every neural architecture apart from the proposed model loses to a three-term linear specification from 2008. Anyone who has tried to improve on HAR for realized variance will recognise the ordering. The paper deserves credit for reporting it plainly.

Three blocks and a learned gate

FinAnalyzeNet, the author's name for the hybrid network, forecasts realized volatility calculated from 5-minute returns. Three blocks supply a gated fusion layer. Feature extraction comes from a temporal convolutional stack with 6 residual blocks, dilations 1 through 32, kernel 3 and 64 filters. Three bidirectional LSTM layers with 256, 128 and 64 units model the sequence. Attention uses Eight heads and key dimension 64. A sigmoid gate learns from the concatenated outputs and controls how much of each component proceeds. The mixture density output then generates 1-day, 5-day and 20-day forecasts with intervals. Training uses a four-term loss covering MSE for point accuracy, QLIKE for the tails, regularisation and a multi-task term. Its weights are learned rather than fixed.

The input set is broad. It includes returns over 1, 5, 10, 20 and 60 days, realized skew and kurtosis, the usual technicals, volume, Amihud illiquidity, bid-ask and order-book measures, FRED and OECD macro series, VIX, put-call ratios, plus BERT-scored news and social sentiment. Recursive elimination selects features on the 2000 to 2017 training block, after which the selection is frozen. That sequence is appropriate.

The sample covers 65 instruments from January 2000 to December 2023, comprising five equity indices, 50 large-cap stocks, four FX pairs and six commodities. There are more than 350,000 daily records. Training runs from 2000-2017, validation from 2018-2020 and testing from 2021-2023. On the test block, FinAnalyzeNet records RMSE of 0.134 versus HAR-RV's 0.152. The reported improvement is 11.8%, with a Diebold-Mariano statistic of 3.42. Directional accuracy reaches 84.7%, compared with 80.3% for HAR-RV and 58.3% for GARCH(1,1). The Model Confidence Set procedure retains models whose forecast losses cannot be statistically distinguished from the best. At 95%, FinAnalyzeNet is the sole survivor.

The paper makes no return forecast. Its economic case rests on downstream uses such as sizing, VaR and option pricing. The economic section therefore carries much of the argument, and it is also where the paper is weakest.

Which gain should readers use?

The multi-scale table gives RMSE improvements over HAR-RV of 11.8% at one day, 12.4% at five days and 15.6% at twenty. For those same horizons, the narrative and conclusion give 10.1%, 13.2% and 15.8%. None agrees with the table, and the discrepancy goes unexplained.

The benchmark count breaks three ways. The abstract says twelve. Support Vector Regression with an RBF kernel appears in the experimental section's model list, while the main results table has eleven rows and omits SVR. The methodology section names thirteen, among them a standalone TCN that appears in neither the setup nor the table. These are small inconsistencies, yet they shape how much confidence a reader places in everything that follows.

The missing units matter more. We did not find any statement in the paper explaining what 0.134 measures, whether daily realized volatility in decimals, annualised volatility or variance. A risk manager cannot translate the 0.018 difference from HAR-RV into an actionable quantity without that information.

Can 84.7% volatility direction hold up?

This result most needs independent examination, and it is the one I trust least. Calling the direction of realized volatility correctly five times out of six on the held-out 2021-2023 block is a substantial claim. GARCH(1,1) reaches 58.3% over the same test block.

The quoted standard deviation across folds is 1.2%, which appears precise until the dates are inspected. Folds 4 and 5 use the identical test window, 2023-01-03 to 2023-12-29. Fold 3 tests from 2022-01-03 to 2023-12-29 and therefore contains both. It also contains Fold 2's complete test period, 2022-01-03 to 2022-12-30, while later validation windows overlap earlier test windows. Dispersion calculated across folds sharing most observations will understate the true variation. The +/-2.3% subsample stability result carries the same weakness.

The paper identifies two causal-validity concerns and addresses them. It masks the backward LSTM pass to the window ending at t-1 and limits attention to time steps before t. Its check remains indirect, relying on the absence of performance jumps at regime boundaries instead of a formal leakage test. The ablation suggests limited stakes either way. A strictly unidirectional version with identical capacity reports RMSE of 0.138 against 0.134, a 3.2% difference, with DM 2.18 and p = 0.015.

The bidirectional pass is a selling point worth 3.2%.

For the regime analysis, expanding-window Bai-Perron break detection recomputes labels using only information available through t-1. The 16.3% high-volatility improvement and 19.1% transitional improvement reported in the abstract come from these expanding-window results. This recomputation is handled properly.

The ablation table contains the paper's strongest internal clue about where performance originates. Removing the TCN raises RMSE by 18.7%, while removing the BiLSTM costs 12.3%. Attention accounts for 8.9%. Removing gated fusion costs 4.2%, a result the author identifies as statistically insignificant with DM 1.92 and p = 0.08 rather than allowing it through unremarked. In the same sentence, he calls the gate essential across diverse market conditions, citing sensitivity evidence that is never displayed. The convolutional feature extractor contributes more than the recurrent core, despite the hybrid recurrent model carrying the headline.

Our adaptation and its limits

We could not reproduce the paper's setup. US minute and daily data on our platform begin around 2010, putting the 23-year window beyond reach. News, transcript and options coverage starts around 2020. We also lack spot FX and FX futures, so we could not attempt the currency-pair validation. High-frequency features had to be assembled from 1-minute OHLCV bars. Tick, quote and order-book inputs were unavailable.

The architecture can be implemented, although the block counts, attention dimensions, TPE search and loss weighting all had to be reconstructed. Complete hyperparameter settings and training details are deferred to a GitHub repository. The paper supplies two different repository URLs: financial_volatility_data in the reproducibility section and FinAnalyzeNet in the data availability statement. Provider terms prevent redistribution of the raw data, leaving a synthetic data generator as the available substitute.

We therefore built an adaptation using a daily volatility-targeting overlay on the top 30 US-listed ETFs by dollar volume. The universe was screened annually and divided into risk-on, Treasury and gold sleeves. A monthly-retrained FinAnalyzeNet-style model forecasts 1, 5 and 20-day realized volatility for each risk-on ETF. We aggregate the 5-day forecasts as a root mean square, then apply a 12% annualized volatility target with an 80th-percentile de-risking overlay. The equity sleeve is equally weighted, with 70/30 Treasury/gold on the residual, a 10% cap per name and cash wherever implementation is infeasible. Trades execute at the closing auction. We charge commissions of $0.0040 per share, subject to a $1.00 order minimum.

Our figures cover our own run from 2020-01-01 to 2024-07-01. Sharpe was 0.64, cumulative return 20.97%, realized volatility 7.28% and maximum drawdown -18.73%, across 5,707 trades. All figures are net of the commissions described above. The paper reports 1.42 for its volatility-based strategy, leaving our 0.64 well below it. Yet the paper gives no universe, period or construction rule for calculating the 1.42. Our result comes from a US ETF overlay of our own design over a window beginning in January 2020.

We can identify the likely direction of each contributor to the gap, though none of their magnitudes. Realized volatility of 7.28% against a 12% target, alongside beta to SPY of 0.18, indicates that the overlay and 10% caps kept the portfolio chronically under-exposed. Returns were compressed without a comparable reduction in drawdown. The result was a Calmar of 0.23 on the -18.73% maximum.

Breadth differs sharply: 30 correlated US ETFs in our run versus 65 instruments spanning four asset classes in the paper. Timing differs too. Our run begins with the March 2020 crash, during which a volatility-targeting rule de-levers and then re-levers late. The paper's test period avoids that drag. Our model also leaves out the training-only correlation filter, RFE-CV selection, the 100-trial TPE search, and the sentiment and microstructure inputs. The defensive sleeve probably detracted as well. The year 2022 represents about a third of our sample, and Treasuries declined alongside equities, so the de-risking leg failed to offset risk as the design assumes.

The visible differences do not fully explain the shortfall. Our outcome is evidence about our implementation before it is evidence about Agal's work.

VaR survives scrutiny; Sharpe remains unsupported

The VaR analysis provides the clearest economic result. Historical simulation using rolling 500-day standardized returns produces a 4.9% violation rate against a 5% nominal rate. Unconditional coverage has p = 0.724 and conditional coverage p = 0.712. GARCH(1,1) records 6.8%, with conditional coverage p = 0.028. This improvement is checkable within a risk application and requires no trading rule.

The 1.42 Sharpe depends entirely on one. It appears only in the abstract and a figure caption. We did not find a description of the trading rule, leverage limits, turnover, transaction cost assumptions, drawdown or the Sharpe from a naive volatility-targeting benchmark. The limitations section openly discusses training cost, reported as 18.7 hours on an A100, along with weaker performance in illiquid markets and incomplete transparency. The trading result receives no such treatment.

Until construction and costs appear on the page, 1.42 is not a finding a desk can price. The 11.8% RMSE advantage over HAR remains the result worth testing.

The paper reports no such comparison. My view would change quickly if the same forecasts were passed through a stated sizing rule with costs charged, turnover disclosed and an otherwise identical HAR-RV-driven version beside it. If the neural forecasts continue to win net, the architecture earns its 18.7 hours. We raised the same objection before about a volatility study whose Sharpe depended on an unstated implementation detail (the 164-year news index). This missing step continues to separate a forecasting paper from a trading paper.

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.

At each calendar-year boundary:
    Select the top 30 ETFs by trailing-dollar-volume screening for that year.
    Classify screened ETFs as risk-on, US Treasury, physical gold/bullion, or ineligible.

At each monthly model refresh:
    Build instrument-level realized-volatility targets from genuine 5-minute returns.
    Make a 1-day, 5-day, or 20-day label eligible only after its full horizon has matured.
    Fit preprocessing, normalization, outlier controls, and any Hansen-Lunde correction
        using expanding training data only.
    Retrain the multi-task volatility model using variable-length sequences ending at
        the information cutoff.

For each trading day t:
    Use only information available through the previous completed session.
    Sequentially classify each instrument's regime from its own volatility history and VIX.
    Forecast each eligible risk-on ETF's 1-, 5-, and 20-day realized volatility.
    Aggregate the 5-day forecasts:
        aggregate_sigma_5d = sqrt(mean(forecast_k_5d^2))
    Feed aggregate_sigma_5d into the intended 12% annualized volatility-target rule,
        subject to the trailing 80th-percentile overlay, 50% high-volatility reduction,
        and maximum leverage of 4.0.
    Equal-weight eligible risk-on ETFs within the resulting equity exposure.
    Allocate residual exposure 70% to eligible Treasury ETFs and 30% to eligible gold ETFs.
    Cap every final ETF weight at 10%; leave infeasible residual weight in cash.
    Skip the rebalance if a required sleeve lacks its minimum eligible instrument count.
    Execute at the current session's official closing auction price; skip any order lacking
        a real daily close and never fabricate an execution price.

After execution:
    Re-price the blotter using the platform's charged commission schedule.