The paper's error conversions fail its own arithmetic. MSE appears on a normalized scale beside RMSE in index points, yet both cannot be right given the printed training range. The abstract pairs MSE of 0.012 on a min-max scale with RMSE of 121.8 index points. Applying the paper's scaling makes those figures differ by a factor of about nineteen. And there is no P&L anywhere in the manuscript.
The front matter carries a warning: this is an unedited accepted version, and "there may be errors present which affect the content." The normalized unit runs through everything that follows, including the ablations, the Diebold-Mariano tests and the claimed 40-52% MSE reduction over baselines at 0.018 to 0.025.
A disclosure comes first. We cannot trade NIFTY-50, so we used liquid US ETFs instead, principally SPY, with QQQ, IWM and sector ETFs optional. Results from that run apply to the adapted US-ETF strategy. They cannot validate the paper's NIFTY-50 error metrics or be transferred to them.
Pimpalkar, Somkunwar, Shelke, Srivastava, Malpe and Kathole built a pipeline with nine stages, four gates and 8 attention heads. Prices undergo min-max normalization using parameters fitted only on the training block. An Isolation Forest then flags days using 100 trees, contamination 0.05 and a threshold at the 95th percentile of training anomaly scores. Those days are masked rather than deleted.
Each rolling window uses the remaining observations to calculate MA20, RSI, MACD, momentum and 20-day rolling volatility. NifSent50, a headline dataset, supplies a daily sentiment score. Feature ranking comes from a gradient boosting model with 100 estimators, lr 0.01 and depth 3. PCA retains 95% of variance while reducing the feature-matrix condition number from 1200 to 45. A Granger screen at max lag 10 examines the surviving features.
The forecast adds two components. ARIMA(5,1,0) handles the linear part, with AIC selecting that order at -2841 versus -2835 and -2838 for the two neighbours. The model is re-estimated on an expanding window before every test day.
QuadGRU handles the remainder. It extends a GRU with two additional gates, forget and context. Each gate reads a binary high-volatility flag defined at the 70th percentile of 20-day rolling volatility. The network has four layers of 128 hidden units and dropout 0.20, then 8-head attention at model dimension 512.
The data consists of NIFTY-50 daily closes from 1 January 2010 to 31 December 2025, with T = 3,928 trading days after cleaning. Tomorrow's close is the target. Chronological splits are 80:10:10, supplemented by seven walk-forward folds. The first trains on 2010-2017 and tests 2019; the last trains on 2010-2023 and tests 2025.
The headline result is 87.5% (±2.1) next-day directional accuracy. Their Encoder Transformer reaches 83.1%, LSTM 79.8% and GRU 78.2%. Against three re-implemented competitors, the Diebold-Mariano statistics range from 3.94 to 5.62.
Any economic value would come from predicting the sign of tomorrow's index close and sizing a position around it. The abstract says the results "provide investors, portfolio managers, and financial analysts with a practical tool to support their investment decisions in volatile markets." The conclusion calls the architecture "making it a dependable decision-making tool for investment strategies and risk management."
Yet the same conclusion identifies the missing trading layer. Future work could turn the model "into a complete automated trading system by combining reinforcement learning-based decision making and execution strategies." This closes a paragraph beginning "Although the results are encouraging, there are some constraints that warrant further investigation." That paragraph also limits the evaluation to a single market and index at daily frequency only. It describes the Granger analysis as "essentially linear", while leaving XAI methods and transformer architectures to later work.
The authors therefore acknowledge that the framework stops short of a trading system, even as the abstract presents it as a practical aid to investment decisions. A forecast-error table does not constitute an investment decision. The paper gives no cost, position sizing or return.
The units do not reconcile
The paper defines min-max scaling and prints its training range: P_min 4,978 and P_max 26,277, spanning 21,299 index points. One normalized unit consequently represents 21,299 points. Using those figures, I calculate that RMSE of 0.110 in normalized space is roughly 2,340 points, or about 11% of the range. The reported value is 121.8.
The same scaling factor, squared, accounts for the MSE discrepancy. Normalized MSE of 0.012 converts to near 5.4 million in squared index points, while the abstract reports 14,850. MAE of 0.084 normalized appears as 93.1 index points, producing the same mismatch a third time. An error of 121.8 points would imply normalized MSE around 0.00003, rather than 0.012.
MAPE offers no resolution. MAE of 93.1 points can align with MAPE of 1.42% only when the average index level in the test sample is around 6,600. The paper's training range reaches 26,277, and its folds test years from 2019 through 2025.
Can 87.5% support the economic claim?
Tomorrow's closing level is the target. The model adds the attention output to the ARIMA forecast, then calculates directional accuracy from changes in the predicted levels. The paper characterizes the closing series as a long-term upward trend broken by 2020 and reports 85.7% accuracy across the full 2010-2025 period. A level-tracking model is therefore scored partly on drift.
No transaction-cost assumption appears. I also did not find a lag-1 persistence or random-walk row in the seven-row comparison table, which contains the proposed model, Encoder Transformer, LSTM, GRU, and Methods [3], [8], [15]. Such a persistence row would show whether 87.5% reflects an edge or follows from the target definition.
Extreme days receive noise treatment as well. Isolation Forest flagged 5.3% of days during 2020, expressly including the March declines, and the reported detection rate for that period is 98.2%. Before ARIMA estimation, flagged days are linearly interpolated. Their attention weights are masked. The regime-conditioned results show high-volatility MSE of 0.0168 versus 0.0084 in low-volatility periods.
The preprocessing section says masking retains sequence length and states, "Observations are marked but not physically removed, preserving sequence length and temporal order." Algorithm 2 step 8 instead says "remove anomalous samples". I did not find a statement clarifying whether the reported errors include flagged test days.
Most of the gain comes earlier
Removing ARIMA raises MSE to 0.0143, while removing attention produces 0.0138, compared with 0.012 for the complete model. Removing sentiment gives 0.0170. Removing the anomaly filter gives 0.0182. The news feature and data cleaning contribute more than the statistical hybrid named in the title.
Directional ablations tell the same story. Without sentiment, directional accuracy falls from 88.2% to 83.0% (t = 4.12, p = 0.003 over 10 runs, 912 test days). The paper also credits that feature with a 9.4% accuracy gain during Q1-Q2 2020. It explains headline preprocessing and daily aggregation, though it states no publication-time cutoff for headlines relative to the close being forecast.
The context gate result deserves attention. Its average activation is 0.73 during low-volatility regimes and 0.31 during high-volatility regimes, p < 0.001. The paper measures this switch directly.
Internal figures remain loosely reconciled, consistent with the front matter's warning that content-affecting errors may remain. Table 5 lists Methods [3], [8] and [15] at 80.3, 78.6 and 74.2%. The conclusion gives 71.0, 80.0 and 84.0%, a ranking that lifts the weakest competitor into second place.
The percentages attached to the ablations also conflict with their MSE ordering: No Sentiment is 0.0170 at 41.70%, whereas No IsFo is 0.0182 at 34.06%. The text says the design "cuts FLOPs by around 51% from the base architecture of GRU". Table 9 reports 187M FLOPs per forward pass, compared with 58-124M for the baselines.
Periods are mixed as well. Table 5 describes its accuracy column as drawn from the 2020 crash and its MSE column as drawn from stable 2018. The abstract assigns 87.5% to 2010-2025 rolling-window validation. A later figure reports 85.7% for the same span.
What our automated pass could measure
Our run uses SPY and optionally QQQ, IWM and sector ETFs because NIFTY-50 is unavailable to us. Any resulting figure tests only the adapted strategy.
An automated pass on our side could quantify three items the paper leaves unreported: a naive-persistence benchmark, given the absence of that row from the comparison table; transaction costs, for which no assumption is stated; and P&L from the sign forecasts, since the paper attaches no economic result to its hit rate. If next-day ETF sign calls beat yesterday's close and always-long after spread, the mechanism would merit further work on our own instruments.