The motivation behind the work stems from a well-known blind spot in classical forecasting. Autoregressive models such as ARIMA and vector autoregression excel at capturing short-term statistical structure in price series, but they treat the market as a closed system. They cannot ingest exogenous information flows such as breaking news, earnings guidance, or macroeconomic events, and they ignore cross-asset structure like sectoral co-movements. When regimes shift, patterns learned from past prices can be invalidated overnight by information that never enters the model. Deep learning models, including long short-term memory networks, improve on this by handling multivariate, nonlinear data, yet most published approaches still analyze a single data source or modality, limiting their effectiveness in highly volatile conditions.

The new architecture addresses these gaps with three parallel branches. The first is a long short-term memory network, a type of recurrent neural network that uses gated memory cells to carry information across many time steps without the vanishing gradient problems that plague conventional recurrent networks. The LSTM branch processes multivariate sequences of stock prices and technical indicators, including exponential moving averages, the moving average convergence/divergence signal, parabolic SAR, Bollinger bands, and stochastic oscillators, all normalized before training. Its job is to extract the temporal fingerprints of each stock’s behavior over recent days.

The second branch is a graph convolutional network, and this is where the model departs most sharply from prior work. Markets are systems of interacting assets: stocks move together in time-varying clusters, and information often diffuses across related companies. Rather than fixing these relationships in advance, as many graph-based approaches do with static industry hierarchies or long-window correlations, the researchers compute a Pearson correlation matrix over the most recent closing-price sequences for all seventeen stocks, add self-loops, and apply symmetric degree normalization. This dynamic adjacency matrix is rebuilt every prediction day, so the graph topology evolves with the market. When correlations strengthen, decay, or invert around a shock, the network’s message passing reflects the new reality rather than a stale snapshot. Each stock node is summarized by seven statistics of its recent series: mean, trimmed mean, standard deviation, trimmed standard deviation, skewness, kurtosis, and trend slope.

The third branch handles language. General-purpose language models often struggle with the specialized terminology and scarcity of labeled data in finance, so the team uses FinBERT, a BERT variant pretrained on the Reuters TRC2-financial corpus of roughly 46,000 documents and fine-tuned on the Financial PhraseBank, a set of 4,845 sentences annotated by finance professionals. FinBERT has been reported to outperform even GPT-4 on financial sentiment benchmarks. For each stock and each day, the model scores every news headline as positive, neutral, or negative, sums the positive and negative scores to determine the day’s overall tone, and then extracts the embedding of the single most representative headline. A separately fine-tuned FinBERT head also produces an uptrend or downtrend signal from the news corpus itself.

Fusion is where the three modalities meet. The embeddings produced by the GCN, LSTM, and FinBERT branches are concatenated and passed through dense layers that refine the joint representation, and a final sigmoid neuron outputs the probability that a given stock will rise the next day. The researchers emphasize that this is more than late concatenation of independent encoders: because the graph branch conditions each asset on its peers, the fused representation is simultaneously context-aware through text and structure-aware through dynamic graphs, while the sequence branch retains sensitivity to temporal dependencies. An ablation study confirmed that removing either the GCN or the FinBERT component degrades performance, with the full model achieving the best average portfolio gains in thirteen of seventeen uptrend cases and ten of seventeen downtrend cases.

The experimental design is unusually demanding. The team evaluated seventeen stocks across five sectors, from technology giants like Apple and NVIDIA to healthcare and real estate firms, over two contrasting periods: an uptrend from July to December 2021 and a severe downtrend from January to September 2022, a stretch the authors note ranks among the worst markets since the 2008 financial crisis. Models were retrained daily in a sliding-window fashion, predicting one day ahead. Against baselines including ARIMA, gradient boosted trees, plain LSTMs, bidirectional LSTMs, and several convolutional sequence-to-sequence architectures, the proposed model delivered competitive F1-scores, leading on stocks such as IBM, NVIDIA, Prologis, and American Tower in at least one market phase. Gradient boosted trees often posted the highest F1-scores, but with an important caveat: they could train on price data reaching back to 2010, while the multimodal model was constrained to the period from 2021 onward, when aligned news data became available.

The most compelling findings emerged from portfolio simulations, where each model’s daily predictions triggered simulated buy and sell decisions starting from ten thousand dollars. Here a disconnect appeared between classification metrics and actual profitability. On Adobe, the proposed model scored lower on F1 than ARIMA and LSTM, yet it delivered average gains of 1.9 percent while ARIMA lost 9.14 percent and LSTM lost 1.62 percent. The authors explain that binary metrics weight every day equally, whereas a model that correctly calls trends on days with large price swings captures genuinely profitable opportunities. Across the full multi-stock swing trading strategy, the proposed model averaged a 7.64 percent gain in the uptrend and a mere 0.34 percent loss in the downtrend, compared with a catastrophic 40.3 percent loss for buy-and-hold and double-digit losses for most baselines. On Apple, it achieved average gains of 24.23 percent in the uptrend and 3.96 percent even in the downtrend, and on Google during the downturn it was the only model to finish positive.

Transparency was a deliberate design goal. The system includes an explainability layer that produces gradient-based feature importance scores for the LSTM branch, time-window relevance plots, correlation heatmaps revealing which stocks most influenced a prediction, and daily sentiment histograms highlighting the most influential news item. These visualizations give analysts a window into why the model expects a rise or fall, rather than presenting an opaque signal. Computationally, the approach is practical: a full cold-start training run of five hundred epochs takes about an hour and a half on a consumer-grade GPU, and daily fine-tuning, with the FinBERT encoder frozen, fits comfortably within a nightly batch window.

The authors are candid about limitations. The model depends on timely and clean news coverage, cannot yet support intraday latency, and its reliance on aligned multimodal data restricts how far back training history can extend. Future work will explore additional modalities, including earnings call transcripts, press releases, and options-implied volatility features such as the term structure of implied volatility and skew, which could provide early warning of regime shifts and crash risk. For now, the study offers a persuasive demonstration that markets cannot be understood through prices alone: the words investors read and the invisible threads connecting companies carry real predictive signal, and a model that listens to all of them at once, while continuously updating its map of who moves whom, may be better prepared for the next storm than anything watching a single chart.

Subject of Research: Multimodal deep learning combining graph convolutional networks, LSTM, and financial language models for stock market trend prediction

Article Title: GCN-FinBERT-LSTM: a graph convolutional deep fusion model for multimodal stock market prediction

Article References: Chen, P., & Corizzo, R. (2026). GCN-FinBERT-LSTM: a graph convolutional deep fusion model for multimodal stock market prediction. International Journal of Data Science and Analytics, 22(1), Article 334. https://doi.org/10.1007/s41060-026-01301-8

Image Credits: AI Generated

DOI: 10.1007/s41060-026-01301-8

Keywords: deep learning, stock market prediction, graph convolutional networks, FinBERT, LSTM, multimodal fusion, sentiment analysis, time series forecasting, financial news, portfolio simulation, dynamic graphs, technical indicators

The framework, called FAGNet, short for Frequency-Guided Adaptive Graph Network, begins with a deceptively simple observation. Financial time series are not single, uniform phenomena. They are superpositions of processes operating at different speeds: fast, high-frequency fluctuations driven by intraday trading and noise, medium-frequency swings tied to weekly and monthly cycles, and slow, low-frequency drifts that reflect macroeconomic fundamentals. When a graph neural network ingests these mixed signals directly, it must somehow learn to disentangle all of these scales simultaneously, a task the authors argue is fundamentally ill-suited to time-domain processing. Their solution is to apply the Fourier transform, a mathematical technique with a century of pedigree in physics and engineering, to project the coupled stock price signals into the frequency domain, where each oscillation speed becomes a distinct, measurable component.

But simply moving to the frequency domain is not enough, and this is where the technical heart of the paper lies. A naive frequency decomposition, such as splitting the spectrum into bands of equal width, would not guarantee that each band carries comparable information. Two bands of identical width can contain wildly different amounts of energy, and therefore wildly different amounts of signal. FAGNet addresses this with an adaptive energy-based spectral partitioning mechanism guided by the power spectral density, a function that describes how the power of a signal is distributed across frequencies. Rather than carving the spectrum into equal slices, the model divides it into sub-bands that each contain roughly equal energy. The result, the authors contend, is genuine scale decoupling at the level of information content: each sub-band represents a comparable share of the market’s total dynamical activity, so no scale is systematically over- or under-represented in the downstream analysis.

Once the spectrum has been partitioned, the framework constructs a dedicated graph neural network for each decoupled sub-band. This is a crucial architectural decision. Risk contagion, the authors argue, is scale-dependent: the network of relationships through which a fast-moving shock propagates may look very different from the network along which slow-moving fundamentals diffuse. A rapid panic might spread through algorithmic trading links and sentiment spillovers, while a slow repricing might follow industrial supply chains and shared analyst coverage, a phenomenon documented in the finance literature. By modelling each frequency band with its own graph network, FAGNet can learn the contagion topology that is specific to that scale, rather than forcing a single graph structure to explain contagion at every speed simultaneously. The separated scale signals thus become the raw material for a family of specialised graph learners, each attuned to one slice of the market’s rhythm.

The final stage of the pipeline tackles a problem that arises whenever multiple models are combined: how should their outputs be fused? A static weighted average would be blind to changing market conditions, yet the relative importance of different scales is precisely what shifts during a crisis. FAGNet therefore employs a cross-scale interaction and context-aware dynamic fusion module. This module does two things. First, it captures synergistic dependencies between risk contagions at different scales, recognising that a slow-moving fundamental weakness can amplify a fast-moving panic, and vice versa. Second, it adaptively adjusts the prediction weight assigned to each scale according to the instantaneous state of the market. In turbulent conditions, the model can lean more heavily on the frequency bands that carry the most relevant information; in calmer periods, it can rebalance toward the slower components. The fusion is not a fixed formula but a learned, state-sensitive mechanism.

The motivation for this frequency-domain perspective draws on a substantial body of prior research. Economists have long known that systematic risk behaves differently at different time scales, with wavelet-based studies of emerging stock markets showing that the relationship between risk and horizon is far from uniform. More recently, the machine learning community has embraced frequency methods for time series: frequency-enhanced decomposed transformers have improved long-term forecasting, frequency-domain multilayer perceptrons have proven to be effective learners, and Fourier-based graph networks have been applied to multivariate prediction problems. FAGNet extends this line of work by combining frequency decomposition with graph-based contagion modelling in a principled way, using the energy content of the spectrum, rather than arbitrary frequency cutoffs, to define the scales themselves.

The empirical case for the approach rests on extensive experiments on real-world stock market datasets, where the authors report that FAGNet significantly outperforms state-of-the-art baseline methods. The baselines against which such models are typically judged include temporal relational ranking systems, dynamic graph attention networks, integrated convolutional and recurrent architectures built on knowledge-incorporated graphs, and hybrid frameworks that combine graph attention with recurrent memory cells. The claim that a frequency-guided design beats these established approaches suggests that the multi-scale entanglement of risk contagion is not merely a theoretical nuisance but a genuine bottleneck limiting the accuracy of existing models. The authors argue that the results validate frequency domain decoupling as a new paradigm for financial risk modelling, one that is both more accurate and more dynamically interpretable than time-domain alternatives.

That word, interpretable, deserves emphasis, because it points to what may be the framework’s most valuable property for practitioners. Deep learning models in finance are often criticised as black boxes: they may predict well, but they offer little insight into why. FAGNet’s architecture is different in kind. Because each graph network operates on a well-defined frequency band, its learned contagion topology can be read as a statement about how risk propagates at that particular scale. The dynamic fusion weights, meanwhile, provide a running record of which scales the model considers most informative at any moment. In principle, an analyst could inspect these weights during a market stress event and see whether the model is relying on fast, panic-like components or slow, fundamental ones. This is a form of transparency that emerges naturally from the design, rather than being bolted on afterwards, and it aligns with a broader movement in the field toward models whose internal structure mirrors the phenomena they describe.

The study also situates itself within a rapidly growing literature on graph-based financial prediction. Graph neural networks have become the dominant approach for modelling risk contagion, with applications ranging from recommending profitable stocks via financial graph attention networks to predicting stock movements using hybrid-relational market knowledge graphs. Yet the authors identify a persistent limitation: the vast majority of deep graph models operate directly in the raw time domain, making it difficult for them to handle the multi-scale entanglement and dynamic variability of contagion. Related work by some of the same authors, including a frequency-domain graph learning framework for understanding risk propagation and a frequency-decoupled progressive graph learning approach for heterogeneous contagion, indicates that this research group has been systematically developing the ideas that culminate in FAGNet. The new paper consolidates that programme into a single adaptive architecture.

Caveats remain, as they always do in financial machine learning. The paper reports that data will be made available on request, and independent replication on different markets, asset classes and time periods will be needed to establish how general the gains are. Markets are adversarial environments where any published predictive edge tends to erode as it is arbitraged away, and no architecture, however elegant, escapes that dynamic. Still, the conceptual contribution stands on its own merits. By treating financial contagion as a multi-scale signal processing problem, partitioning the spectrum by energy content, and letting each scale have its own graph model with state-aware fusion, FAGNet offers a template that could plausibly extend beyond equities to other networked dynamical systems, from cryptocurrency markets to interbank lending networks. For a field searching for ways to make deep learning both sharper and more transparent, the message is that sometimes the answer is not a bigger network, but a better coordinate system in which to run it.

Subject of Research: Frequency-domain graph neural networks for financial risk contagion and stock market prediction

Article Title: Fagnet: an adaptive frequency-guided multi-scale graph network for financial market prediction

Article References: Zhou, C., Xiao, S., & Hu, C. (2026). Fagnet: an adaptive frequency-guided multi-scale graph network for financial market prediction. International Journal of Machine Learning and Cybernetics, 17(10), Article 485. https://doi.org/10.1007/s13042-026-03301-3

Image Credits: AI Generated

DOI: 10.1007/s13042-026-03301-3

Keywords: graph neural networks, stock market prediction, frequency domain analysis, risk contagion, Fourier transform, power spectral density, multi-scale learning, financial networks, deep learning, dynamic fusion, interpretable AI, machine learning

The study addresses a problem that has repeatedly limited the usefulness of artificial intelligence in finance: markets do not behave like fixed systems. Many prediction models process stocks as time series, examining historical prices, trading volumes, volatility, and technical indicators. Others represent companies as nodes in a graph, connecting them through sector membership, correlations, ownership structures, or shared information. These methods can be powerful, but they often treat relationships as symmetrical. If the movement of one asset is associated with the movement of another, a conventional model may effectively regard the connection as operating in both directions. Financial influence, however, is rarely so balanced. A large bank may affect smaller financial firms more strongly than those firms affect the bank; a dominant chip manufacturer may influence equipment suppliers more than the reverse. AsymAlpha was created to reflect this directional and unequal flow of information.

At the heart of the framework is the Dynamic Asymmetric Relationship Module, or DARM. This component uses an asymmetric attention mechanism to estimate the strength of directional relationships among assets. In attention-based neural networks, the model assigns different weights to different pieces of information, allowing it to focus on the signals most relevant to a particular prediction. DARM extends that principle from individual features or time steps to relationships between companies. For one stock, it can learn that the recent behavior of another asset deserves considerable weight; for a different stock, the same relationship may be weak or irrelevant. Because the attention is asymmetric, the influence assigned from asset A to asset B need not equal the influence assigned from B to A. The resulting representation is intended to capture a market information-flow graph rather than a simple network of undirected correlations.

The researchers also impose structural guidance through a differentiable Directed Acyclic Graph constraint. This design draws on continuous optimization methods for causal structure learning, including the “DAGs with No Tears” approach cited by the authors. A directed acyclic graph is a network of arrows that contains no closed loops: information may flow from one node to another, but it cannot return to its starting point through a circular chain. In practical terms, the constraint encourages the model to learn an organized directional structure instead of producing an unrestricted collection of potentially unstable connections. The constraint is differentiable, meaning it can be incorporated into gradient-based training rather than applied only after the neural network has been fitted. This allows the relationship-learning process and the forecasting objective to influence one another during optimization. The resulting graph should not automatically be interpreted as definitive economic causality, but it gives the model a more structured way to represent directional dependencies.

A second component, the Market-Gated Predictor, or MGP, is intended to solve another weakness of financial forecasting systems: the same relationship may matter differently in different market regimes. Markets can shift from calm, low-volatility trading to panic, recovery, momentum, or broad risk-off conditions. In one regime, sector relationships may dominate; in another, macroeconomic or liquidity-related effects may overwhelm company-specific signals. MGP uses context states derived from market snapshot data to regulate the forecasting process. These context states act as a learned gate, helping the model decide which information and relational patterns should receive greater emphasis at a given moment. Instead of applying a single static prediction rule across an entire historical period, AsymAlpha attempts to adapt its internal reasoning to the surrounding market environment.

The complete system is trained end to end, linking temporal market representations, dynamic asset relationships, and the final movement prediction in one learning process. This matters because separately constructed pipelines can create mismatches: a graph may be estimated using one objective, while the predictor is trained using another. AsymAlpha instead allows errors in the final forecasting task to influence how relationships are learned. Its design combines ideas from several active areas of machine learning, including attention mechanisms, graph representation learning, causal discovery, regime-sensitive modeling, and time-series prediction. The framework is aimed at stock movement classification, which generally means estimating whether an asset’s price will rise or fall over a defined future interval. Such predictions can then be translated into portfolio decisions in a simulated trading environment, although the gap between a successful backtest and reliable live trading remains substantial.

According to the article, experiments were carried out on benchmark datasets spanning four major international markets. The authors state that AsymAlpha significantly surpassed competing methods in both prediction accuracy and simulated trading results. The comparison is positioned against a broad range of approaches, including recurrent neural networks, attention-based models, graph-based forecasting systems, market-guided transformers, and methods that use causal or relational information. The reported advantage suggests that learning directional relationships dynamically may offer value beyond simply adding more historical features. It also supports the idea that market context can help a model decide when a learned connection is useful. However, the article preview does not provide the detailed numerical results, dataset names, transaction-cost assumptions, trading periods, or statistical significance tests. Those details are essential for independently judging the size and robustness of the claimed improvement.

The promise of the framework lies not in the fantasy of perfect prediction, but in its attempt to make financial models less rigid. Traditional statistical tools such as autoregressive models often struggle with nonlinear relationships, changing volatility, and structural breaks. Deep neural networks can learn nonlinear patterns, yet they may also memorize historical regularities that disappear when market conditions change. Static graphs face a related problem: correlations and influence pathways observed during one period may weaken or reverse during another. By combining a changing relational graph with a market-state gate, AsymAlpha directly targets this instability. The approach could potentially be extended beyond equities to exchange-traded funds, commodities, cryptocurrencies, or cross-market systems in which information travels across asset classes. Its architecture may also provide researchers with visualizable relational structures that can be studied alongside traditional economic indicators.

There are, nevertheless, important reasons to interpret the findings carefully. A model’s attention weights or learned arrows do not by themselves prove that one company causes another to move. Hidden variables, simultaneous reactions to news, market-wide shocks, and data-processing choices can all create apparent directionality. A differentiable DAG constraint imposes mathematical organization, but it cannot guarantee that the learned graph corresponds to a true real-world causal mechanism. Simulated trading performance can likewise be sensitive to portfolio construction, rebalancing frequency, leverage, short-selling rules, liquidity assumptions, and transaction costs. Financial datasets are especially vulnerable to look-ahead bias, survivorship bias, and changing market composition. The strongest evaluation would therefore require long out-of-sample periods, realistic execution assumptions, repeated tests across market regimes, and transparent ablation studies showing how much each AsymAlpha component contributes.

Even with those caveats, the study reflects a significant direction in the evolution of financial artificial intelligence. The most sophisticated forecasting systems are moving away from viewing assets as isolated price histories and toward treating markets as adaptive information networks. AsymAlpha’s central claim is that these networks should be directional, unequal, and sensitive to changing conditions. Its DARM module seeks to uncover the evolving pathways through which assets influence one another, while its MGP module attempts to recognize the market state in which those pathways become meaningful. If future research confirms the reported gains under independent and realistic testing, such systems could help analysts build more responsive tools for market monitoring and risk assessment. They will not eliminate uncertainty from investing, but they may offer a more nuanced computational lens on the complex forces that make markets move.

Subject of Research: Dynamic, asymmetric, and market-state-adaptive relational learning for stock price movement prediction

Article Title: Dynamic asymmetric relational learning for stock price movement prediction

Article References: Yang, R., Fan, M., Mo, F. et al. “Dynamic asymmetric relational learning for stock price movement prediction.” Data Mining and Knowledge Discovery, 40, Article 41 (2026). https://doi.org/10.1007/s10618-026-01201-2

Image Credits: AI Generated

DOI: 10.1007/s10618-026-01201-2

Keywords: Stock movement prediction, market modeling, algorithmic trading, asymmetric attention, dynamic relational learning, directed graphs, causal structure learning, financial forecasting