The paper earns attention through the gap between two ablation rows: 39.87 against 30.22. Its next-day index-level forecast, with a MAPE of 0.59%, says little by itself about whether anyone can trade it.
Bohra and Vijay call the model VertiFuseX. Three recurrent branches process the same window of past closing prices in parallel, with fusion at the penultimate layer rather than the output. One branch is a 2-layer LSTM with 128 then 64 units. Another is a 2-layer Bi-LSTM with 128 then 32 per direction. The third is a 3-layer stacked LSTM arranged as 128/128/64. Alongside them, a parallel MLP consisting of flatten, 128, 64 reads the window without recurrence.
Each recurrent branch yields a 64-dimensional vector from its last hidden layer, the tensor available immediately before the branch's regression head compresses it to a scalar. VertiFuseX concatenates and reweights the three recurrent vectors first. A second learned layer combines that output with the aligned MLP vector, and a linear head produces the one-step-ahead price. The whole model trains jointly under MSE.
Fusion location is the paper's central claim. Late fusion averages three scalar forecasts after each regression head has discarded information. Combining the representations before that compression could retain more of the input structure. The paper derives the corresponding mutual-information inequality, then limits the claim to a representational property under a sufficiency assumption about the fusion head. Its role is empirical motivation rather than a guarantee.
Then comes the test.
The broad headline reports MAPE reductions of 30 to 54%, along with over 40% improvements in MAE and RMSE, against LSTM-family baselines. VertiFuseX also beats seven published models across 33 metric-by-dataset comparisons on ten indices. One fixed hyperparameter configuration covers the lot.
For the baseline comparisons, the authors use Yahoo Finance daily closes from 2010 to 2024 for S&P 500, DJI, NYSE and NASDAQ. Each index begins with 3,773 raw observations and retains 3,753 after applying the 20-day window. The chronological split assigns 3,388 days to training and reserves the final 365-day test year. The Min-Max scaler is fitted only on the training partition, and non-trading days are dropped instead of forward-filled.
Every index uses dropout 0.30, l2 1e-4, Adam at 1e-4, batch 64, 50 epochs and seed 150. VertiFuseX contains 675,716 parameters, of which 673,924 are trainable, and weighs 2.58 MB. Per-sample inference takes 1.47 to 1.56 ms. For a second comparison set, the authors retrain VertiFuseX on the exact index and period used by each of the seven published models, covering 1991 to 2022.
We could not run the model. Its architecture requires training compute we do not have, so none of what follows is a backtest of their method. Any version we build from the idea would be a substitute, most likely one of their own baselines, and would leave their claim untested. Most of the paper's international indices are also unavailable as holdable instruments. An implementable version would use SPY, QQQ, IWM and sector ETFs, which substitutes a different market. The mechanism is a lagged forecast of the price path rather than anything market-specific, so it ports.
The ablation is worth stealing
On the S&P 500 test year, the standalone stacked LSTM (St-LSTM) records MAE 46.49. Averaging scalar predictions from the three branches lowers it to 43.18, about 7% better. Concatenating the final-layer hidden states and fitting a shared head reaches 39.87. Penultimate fusion brings MAE down again to 30.22, a further 24.2% below the final-layer variant. NASDAQ follows the same order, moving from MAE 170.42 to 134.54 and from MAPE 1.32% to 0.83%.
The authors say the two fusion variants have comparable parameter counts at the combination stage. Their branch architectures are identical across the two rows. Only the tensor selected for concatenation changes.
A confound remains in the average-ensemble comparison. Its branches train independently, which means the move from 43.18 to 39.87 combines learned weighting with joint optimisation.
Using squared-error loss and the Harvey-Leybourne-Newbold correction, the Diebold-Mariano tests produce statistics of 2.68 to 2.96 against St-LSTM, the strongest single branch. Across all four indices, p-values range from 0.0033 to 0.0077. Against ARIMA, the statistics span 3.58 to 4.18. Each test has 365 observations.
I could not find a naive persistence baseline. For a next-day index-level forecast, MAPE around 0.5 to 0.8% lies close to the result from predicting yesterday's close. ARIMA appears in the table and loses, though ARIMA with AIC-selected orders differs from y_{t+1} = y_t. The missing row leaves unresolved how much of the gain comes from level-tracking rather than change-prediction. Saliency raises the same question.
Saliency peaks in the wrong place
Across the 20-day window, gradient saliency peaks at lags 9 to 15. Its location varies by index: NASDAQ 10 to 12, NYSE 9 to 11, S&P 500 12 to 14 and DJI 14 to 16. A next-day price-level forecast should lean heavily on the latest observation because that observation supplies most of the answer.
The paper interprets the pattern as evidence that the fusion layer suppresses high-variance recent lags and favours mid-range context. Another reading is that the network learned a smoothed level estimate instead of a change estimate. Once the level is implicit, recent lags carry less gradient. Persistence would distinguish those explanations, which makes it row zero for any replication.
What does the 2024 simulation establish?
The trading rule enters long when the predicted price change exceeds the 25th percentile of the forecast distribution. A position exits at a +6% take-profit, a -3% stop-loss, or when the signal ceases after a minimum three-day hold. Any stop-loss exit triggers a three-day cool-down. The simulation charges 10bp proportional per trade, executes at the next-day open, uses no leverage, and holds only long or cash positions.
The authors describe this exercise as "simplified trading logic and conservative assumptions, serving primarily as a proof of concept rather than a fully optimized trading system." Yet its threshold is calculated from the test-period forecast distribution. The paper is explicit: "a long position is initiated only when the predicted price change exceeds the 25th percentile of the test-period forecast distribution." That quantile is unavailable in advance. In the stress test, the authors instead use the 75th percentile of the training-period distribution, which is causal. The experiments therefore apply different rules.
Net of 10bp, the 2024 S&P 500 strategy returns 29.41%, with Sharpe 1.43 and max drawdown -9.75%. Buy-and-hold delivers 29.44%, Sharpe 1.37 and max DD -10.28%. Across 31 trades, the strategy roughly matches the index while remaining in cash about 25% of the time.
DJIA returns 21.44% with Sharpe 1.15, compared with 21.08% and 1.06. NASDAQ produces 26.59% and Sharpe 1.04, trailing buy-and-hold at 36.25% and Sharpe 1.27, although max drawdown improves from -13.15% to -11.97%. NYSE fares worse: 7.68% with Sharpe 0.35 against 17.31% and Sharpe 0.88. Its drawdown also deteriorates, reaching -13.14% against -10.66%.
The authors report that their strategy loses to buy-and-hold on two of four indices. They label NYSE "the weakest case in the study." Their conclusion is carefully limited: the results "support the signal-level economic relevance of VertiFuseX, but they do not establish uniform dominance over passive investment." The paper makes a restrained economic claim, and the evidence warrants that reading.
Four correlated US indices during one strongly bullish year provide little information. The threshold looks ahead, and the paper runs no significance test on any Sharpe difference. A long-or-cash rule invested 74.8% of the time during a year when the S&P returned 29% will tend to look acceptable almost regardless of forecast quality.
The S&P 500 stress tests carry more weight. During the COVID crash from 19 Feb to 30 Apr 2020, the model is frozen after training on 2010 to 2019. It returns -9.63% against buy-and-hold's -16.08%, while max drawdown is -26.45% against -33.67%. During the 2022 bear market, the strategy returns -10.04% against -19.56%, with max DD of -17.67% against -25.38%. These runs use frozen parameters, genuinely unseen regimes and the causal training-period threshold. They provide the paper's strongest economic evidence.
One model gets the seed interval
Three seeds (42, 101, 150) produce S&P 500 MAE 30.35 ± 1.69 and NASDAQ 134.77 ± 7.30, using a t-multiplier of 4.303. Across the paper's four baseline indices, every upper bound remains below the best single-branch baseline. Those intervals cover VertiFuseX alone. The comparison therefore places a three-seed interval for VertiFuseX beside single-seed point estimates for LSTM, Bi-LSTM and St-LSTM.
The issue becomes sharper in the seven-model comparison, where error figures come from the original publications rather than reruns under matched splits and seeds. Against NuNet on the S&P 500, VertiFuseX leads by 3.1% MAE, at 12.21 against 12.60. Against reservoir computing on the Nikkei 225, the advantage is 1.83%, with 157.82 against 160.77.
Set those margins beside the seed variation. The ±7.30 around a mean of 134.77 is about ±5.4%, wider than either advantage. Results on the Asian indices are much larger. HSI MAE is 34.63% lower, at 268.84 against 411.27. KOSPI improves by 28.2% on MAE and 27.2% on MAPE. Those gaps can absorb the observed noise. Margins of 2 to 3% cannot, and treating each as a win overstates the 33-comparison count.
In the paper's account, hyperparameters came from standard practice in financial time-series forecasting and preliminary stability checks across multiple indices. They were then fixed. The final 365 days were excluded from training, early stopping and the learning-rate schedule. The authors do not identify which indices were used for the preliminary checks, nor do they name a set disjoint from the ten later evaluated. The fixed configuration is therefore a design choice informed by prior exposure to these markets, rather than a leak of test observations.
The paper acknowledges that it "does not yet employ a full rolling-window walk-forward protocol with multiple overlapping folds," and promises one in future work. Publication proceeded with the single-window result. I would push back there.
The Bi-LSTM processes lags t-20 to t-1 in reverse, but every lag still precedes the forecast origin. The look-ahead objection fails in this case, as the authors argue. The branch does require the complete window to be buffered. Daily data makes that easy; higher frequency turns it into a constraint.
Penultimate fusion beating final-layer fusion by 24.2% MAE would justify the compute once it survives a training-period threshold and multiple folds.