💰 quantitative finance

VertiFuseX: Generalizable Financial Forecasting via Multi-Stream Temporal Fusion

VertiFuseX is a lightweight, hybrid LSTM architecture that achieves superior generalization and accuracy in financial forecasting by employing a novel penultimate-layer vertical fusion of multi-scale temporal representations, significantly outperforming existing baselines across diverse global markets while maintaining deployment-ready efficiency.

Original authors: Aashish Bohra, Vivek Vijay

Published 2026-09-14✓ Author reviewed ⓘ

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Aashish Bohra, Vivek Vijay

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Predicting the future price of a stock is one of the most difficult challenges in finance. The market is not a steady, predictable machine; it is a chaotic, noisy system that changes constantly. Prices do not move in straight lines but jump up and down based on a mix of immediate news, long-term trends, and the shifting moods of millions of investors. Because the data is so messy and changes so quickly, standard computer programs often struggle to find the true signal hidden within the noise. They tend to either memorize past mistakes or get confused when the market suddenly shifts into a new pattern. For decades, researchers have tried to build better tools, using complex mathematical models to guess where prices will go next, hoping to turn those guesses into reliable forecasts that can help investors manage risk.

A team of researchers at the Indian Institute of Technology Jodhpur has proposed a new approach called VertiFuseX to solve this problem. Instead of trying to force the computer to learn a single, rigid way of looking at the past, they built a system that listens to three different "voices" of the market simultaneously. Imagine trying to understand a complex story by listening to three different people: one who remembers the exact order of events, one who understands the context of the whole story from beginning to end, and one who sees the big picture of how the plot has evolved over time. The researchers combined these three perspectives into a single, unified model. They found that by merging these different views of the past before making a final prediction, the system could see patterns that the others missed, leading to significantly more accurate forecasts.

The core of their discovery lies in how they combined these different perspectives. In many previous attempts, researchers would let each computer model make its own guess and then simply average the results, like taking a vote. The new study argues that this method throws away too much valuable information. Instead, the researchers decided to combine the internal "thoughts" of the models while they were still processing the data, before they had to commit to a final number. They took the intermediate representations from their three different models—one that tracks short-term memory, one that looks at the past and future within a specific window, and one that understands deep, layered trends—and fused them together. This allowed the system to keep the rich, detailed texture of the market's history, rather than reducing it to a single, compressed guess.

To test if this idea worked, the researchers trained their system on fifteen years of daily closing prices from ten major stock markets around the world, including the United States, Europe, and Asia. They used a strict testing method where the model was trained on the past and then asked to predict the future without ever seeing the test data beforehand, ensuring the results were not just lucky guesses. The results were striking. The new system reduced prediction errors by between 30 and 54 percent compared to older, standard models. It performed consistently well across different types of markets, from the stable, slow-moving blue-chip stocks to the highly volatile technology-heavy markets that change rapidly. The system was also surprisingly efficient, requiring very little computer memory and making predictions in less than two milliseconds, which is fast enough for real-time trading.

The researchers also looked at how the system made its decisions to understand what it was actually learning. By analyzing which days in the past the model paid the most attention to, they discovered that it consistently focused on a specific window of time, roughly nine to fifteen days ago. This suggests that the model found a "sweet spot" in the data where the signal was strong enough to be useful but not so old that it was irrelevant. This insight is valuable because it shows the system learned to ignore the immediate, chaotic noise of the last day or two and the distant, faded patterns of months ago, focusing instead on the intermediate trends that often drive price movements.

To see if these better predictions actually mattered in the real world, the team simulated a trading strategy using the model's forecasts. They tested how the system would have performed during two of the most stressful periods in recent financial history: the market crash caused by the pandemic in 2020 and the bear market of 2022. In both cases, the system managed to protect capital better than simply holding onto stocks. It reduced the maximum amount of money lost during the downturns and achieved better returns relative to the risk taken. While it did not always beat the market in total profit, it did so with less volatility and fewer sharp drops, demonstrating that the model could adapt to extreme conditions without breaking down.

The study concludes that this new method of fusing different types of memory at an intermediate stage offers a powerful way to handle the complexity of financial markets. It suggests that the key to better forecasting is not just making the models deeper or more complicated, but rather teaching them how to share and combine their different ways of seeing the past. The system proved to be lightweight, fast, and robust across different countries and market conditions, offering a practical tool that could be deployed in real-world financial systems. While the researchers acknowledge that no model can predict the future with perfect certainty, especially during sudden global shocks, their work provides a clearer, more reliable way to navigate the uncertain waters of the stock market.

Technical Summary: VertiFuseX

Problem Statement

Accurate stock price prediction remains a formidable challenge due to the non-stationary, high-dimensional, and noisy nature of financial time series. Existing deep learning (DL) models, including standard LSTMs, Bi-LSTMs, and hybrid architectures (e.g., CNN-LSTM), face three systemic limitations:

Rigid Decision-Level Fusion: Many models collapse multi-scale temporal dependencies into scalar predictions before integration (e.g., via static averaging), preventing meaningful interaction between heterogeneous dynamics like short-term volatility and long-term trends.
Overfitting and Tuning Sensitivity: Deep ensemble models often suffer from generalization collapse due to excessive architectural depth, lack of regularization, and reliance on ad hoc, dataset-specific hyperparameter tuning.
Single-Scale Temporal Modeling: Most LSTM-based approaches capture patterns at a fixed resolution, failing to disentangle hierarchical dependencies across different time horizons (e.g., intraday noise vs. monthly trends), which is critical in volatile markets.

These issues lead to information loss, overfitting, and limited cross-market generalization, particularly when models are tested on strict chronological out-of-sample data.

Methodology: VertiFuseX

The authors propose VertiFuseX, a hybrid LSTM architecture designed to address these limitations through penultimate-layer vertical fusion and multi-scale temporal abstraction.

Architecture Overview

The model operates in three stages:

DATAETL (Data Extraction, Transformation, Load):
- Uses strictly univariate input (raw closing prices) to isolate the architectural contribution from feature engineering.
- Employs a 20-day chronological moving window.
- Implements rigorous preprocessing: non-trading days are removed (not interpolated) to preserve signal fidelity, and Min-Max normalization is fitted only on the training partition to prevent look-ahead bias.
- Data is split chronologically: ~14 years for training/validation and a strict 365-day held-out test set.
FeatExt (Feature Extraction):
- Utilizes three parallel temporal branches, each extracting features from the penultimate layer (before task-specific regression) to preserve rich intermediate representations:
  - LSTM: Captures sequential memory (unidirectional).
  - Bi-LSTM: Captures bidirectional context within the lookback window.
  - St-LSTM (Stacked): Captures hierarchical trends via progressive abstraction across layers.
- A parallel DNN stream processes the same input window to capture non-recurrent, static non-linear relationships.
- All branches are regularized with dropout (0.3) and $\ell_2$ constraints, and outputs are aligned to a shared latent dimension ( $d_T = 64$ ).
VertiFuse (Vertical Fusion):
- Instead of averaging final scalar predictions (decision-level fusion), VertiFuseX concatenates the penultimate feature vectors from the LSTM, Bi-LSTM, St-LSTM, and DNN branches.
- These concatenated features undergo a learned affine transformation (via trainable weight matrices $W_{fuse}$ and $W_{comb}$ ) followed by a non-linearity (ReLU).
- This feature-level fusion allows the model to dynamically reweight complementary temporal representations (e.g., emphasizing Bi-LSTM during high noise or St-LSTM during stable trends) before the final regression head.
- The entire system is jointly optimized via backpropagation under a fixed hyperparameter configuration, eliminating the need for per-dataset tuning.

Theoretical Motivation

The core hypothesis is that penultimate-layer representations retain higher mutual information with the input signal ( $I(X; F_i) \geq I(X; \hat{y}_i)$ ) compared to compressed final outputs. By fusing at this intermediate stage, the model preserves the unique inductive biases of each branch (e.g., bidirectionality, hierarchy) and enables joint gradient propagation, fostering complementary specialization rather than averaging already-degraded predictions.

Key Contributions

Penultimate-Layer Vertical Fusion: Introduction of a novel fusion strategy that integrates intermediate latent features from heterogeneous temporal models (LSTM, Bi-LSTM, St-LSTM) rather than final outputs. This preserves multi-scale temporal abstractions and facilitates feature-level interaction.
Robust Cross-Market Generalization: Evaluation across 15 years of data (2010–2024) on 10 global equity indices (including S&P 500, NASDAQ, HSI, KOSPI) using a strict chronological out-of-sample protocol. The model achieves consistent performance without index-specific hyperparameter tuning.
Elimination of Ad Hoc Tuning: The use of a fixed, pre-validated hyperparameter configuration (window size=20, dropout=0.3, $\ell_2=10^{-4}$ ) across all datasets reduces overfitting and tuning overhead.
Lightweight and Efficient: The model is compact (675k parameters, ~~2.6 MB footprint) with low inference latency (~~1.5 ms/sample), making it suitable for real-time deployment.

Experimental Results

The model was evaluated against standard baselines (LSTM, Bi-LSTM, St-LSTM, ARIMA) and seven state-of-the-art models (e.g., ModAugNet, Reservoir Computing, GA-CNN-LSTM).

Performance Metrics: VertiFuseX achieved 30–54% reductions in MAPE and over 40% improvements in MAE and RMSE relative to LSTM-based baselines.
State-of-the-Art Comparison: It consistently outperformed seven SOTA models across 33 metric-dataset comparisons, with reductions in MAE up to 34.6% and RMSE up to 36.4% compared to the best reported results in prior literature.
Ablation Studies: Experiments confirmed that penultimate-layer fusion drives these gains, yielding substantial improvements over both final-layer fusion and decision-level ensembling (Average Ensemble).
Saliency Analysis: Gradient-based saliency maps revealed a consistent emphasis on mid-range temporal dependencies (lags 9–15 days), suggesting the model learns to balance noise reduction with trend characterization without predefined indicators.
Economic Validation: Algorithmic trading simulations (including stress tests during the 2020 COVID-19 crash and 2022 bear market) demonstrated reduced maximum drawdowns and superior risk-adjusted returns (Sharpe/Sortino ratios) compared to Buy-and-Hold and Momentum benchmarks, particularly in volatile regimes.

Significance and Claims

The paper claims that VertiFuseX advances financial forecasting by unifying simplicity, interpretability, and cross-market robustness. Its significance lies in demonstrating that:

Feature-level fusion at the penultimate layer is superior to decision-level ensembling for preserving temporal structure in non-stationary data.
Multi-scale abstraction via specialized recurrent branches (LSTM, Bi-LSTM, St-LSTM) can be effectively integrated without complex attention mechanisms or heavy hyperparameter tuning.
The architecture offers a deployment-ready framework that balances high predictive accuracy with low computational cost, addressing the practical needs of real-time financial systems.

The authors explicitly note that while the model shows strong robustness, it relies on univariate price data and fixed hyperparameters, representing a trade-off between peak per-market accuracy and broad generalization. The work does not claim to solve all market prediction challenges but provides a validated architectural intervention for improving temporal modeling fidelity in hybrid deep learning systems.