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, named TED-FinRisk, was developed by Wenchuan Kuang and Yusiyuan Chen of Fudan University’s College of Computer Science and Artificial Intelligence, together with colleagues from Gientech Technology’s Financial Risk Lab, Mashang Consumer Finance’s AI Research Lab, and Pennsylvania State University. Its central claim is methodological: a risk taxonomy built from real case texts can carry genuine regulatory semantics if data-driven topic discovery is deliberately calibrated by domain experts rather than left to operate unsupervised. The researchers tested this idea on a corpus of more than 1,000 illegal fundraising cases and nearly 6,000 associated risk-related records, a body of material rich enough to capture the operational texture of contemporary schemes.

Technically, TED-FinRisk is a hybrid in the strict sense, chaining several well-established text analytics techniques into a sequence in which each stage constrains the next. The pipeline begins with term frequency–inverse document frequency, or TF-IDF, feature construction. TF-IDF assigns weights to words based on how often they appear in a given case document and how rare they are across the whole corpus, so that boilerplate legal language fades into the background while distinctive vocabulary, the telltale phrases of a Ponzi pitch or a crypto yield promise, rises to the surface. These weighted vectors become the raw material for everything that follows.

Before any topics are formally extracted, the researchers employ t-distributed stochastic neighbor embedding, or t-SNE, to assist topic exploration. t-SNE is a nonlinear dimensionality reduction technique that projects high-dimensional document vectors into a low-dimensional map while preserving local neighborhoods, allowing analysts to see clusters of similar cases at a glance. In TED-FinRisk this step serves as a visual diagnostic: it helps the team gauge how many natural groupings exist in the corpus and whether candidate topic counts are plausible before committing to a formal model. It is a pragmatic use of visualization as a sanity check rather than an end in itself.

The formal topic extraction is performed with latent Dirichlet allocation, or LDA, a generative probabilistic model that treats each document as a mixture of hidden topics and each topic as a probability distribution over words. Running LDA over the TF-IDF-weighted case texts yields candidate topics, each characterized by its most probable terms, which the researchers can then read as proto-categories of illegal fundraising behavior. But LDA output alone is notoriously noisy and does not automatically align with the categories regulators actually use. This is where the framework’s distinguishing move comes in: expert calibration through the Analytic Hierarchy Process, or AHP.

AHP is a structured decision-making method in which experts compare criteria pairwise and derive consistent priority weights. In TED-FinRisk, the AHP stage sits inside a threat–vulnerability–consequence perspective borrowed from the Financial Action Task Force, the intergovernmental body that sets global anti-money-laundering standards. Experts evaluate the data-driven topics against FATF-style considerations of threat, vulnerability, and consequence, assigning priorities that determine how the candidate topics are merged, split, renamed, and organized into a final taxonomy. The result is a classification scheme with six primary categories and a set of secondary labels, each linked to operational indicators that can feed downstream monitoring systems.

The scope of the corpus gives the taxonomy its breadth. The cases span scenarios ranging from elder care schemes that exploit aging populations’ retirement savings, to virtual currency offerings riding waves of speculative enthusiasm, to supply-chain finance arrangements and agricultural ventures that wrap illicit fundraising in the language of legitimate business. Film investment products, social e-commerce schemes, and overseas investment vehicles round out the collection. This diversity matters because the study’s core argument is that a taxonomy derived from a narrow slice of cases will fail the moment schemes evolve beyond it, whereas a taxonomy grounded in thousands of records across many domains has a better chance of anticipating the next mutation.

Crucially, the authors do not stop at building the taxonomy; they test whether it corresponds to something real in the text. Once the expert-finalized labels are in place, the team runs a reclassification consistency check using FinBERT, a transformer language model pre-trained on financial text. FinBERT is tasked with assigning case documents to the taxonomy’s categories, and the researchers examine whether the model’s assignments align with the expert labels. If a taxonomy category cannot be recovered from the raw text by a semantic model, that category may be an artifact of expert convention rather than a pattern with genuine textual signature. The consistency check thus functions as an empirical bridge between human-defined regulatory semantics and machine-detectable language patterns, a form of expert-in-the-loop validation that the authors present as a reusable design principle.

The paper is explicit about its own character: it is a methodology-oriented case study, not a deployed production system. Its contributions are framed as a reusable taxonomy design process, a structured basis for future benchmarking, and a foundation for early-warning applications. That framing is honest about the state of the field. Static rule lists and purely expert-driven typologies, the authors note in their abstract, have become difficult to maintain as illegal fundraising migrates to dispersed digital channels and rapidly changing narratives. What they propose instead is a repeatable pipeline in which topic discovery, expert prioritization, and semantic validation reinforce one another, so that the taxonomy can be refreshed as new case texts arrive rather than rewritten from scratch.

For regulators and financial institutions, the significance of TED-FinRisk lies less in any single algorithm than in the architecture of collaboration it demonstrates. TF-IDF, t-SNE, LDA, AHP, and FinBERT are all established tools; the innovation is the disciplined order in which they are combined and the insistence that neither the data nor the experts have the final word alone. The six-category taxonomy with its operational indicators offers a template for how machine learning and regulatory practice can meet in the middle, producing classifications that are simultaneously discoverable in the data, meaningful to compliance officers, and testable by independent models. As financial crime continues to evolve at the speed of digital marketing, frameworks of this hybrid kind may become essential infrastructure for keeping watchlists, warning systems, and enforcement priorities a step ahead of the schemes they are designed to catch. The study, published open access on 3 October 2026 and supported by China’s National Key Research and Development Program, the National Natural Science Foundation of China, and regional science projects, invites other jurisdictions to replicate the process on their own case corpora and compare the resulting taxonomies, turning what has long been an artisanal exercise in expert judgment into a benchmarkable, evolving science.

Subject of Research: Machine learning-based risk taxonomy construction for typologizing illegal fundraising from multi-source case texts

Article Title: TED-FinRisk: a hybrid topic-enhanced framework for typologizing illegal fundraising risk from multi-source case texts

Article References: TED-FinRisk: a hybrid topic-enhanced framework for typologizing illegal fundraising risk from multi-source case texts. (n.d.). https://doi.org/10.1186/s40537-026-01574-7

Image Credits: AI Generated

DOI: 10.1186/s40537-026-01574-7

Keywords: illegal fundraising, risk taxonomy, topic modeling, LDA, FinBERT, TF-IDF, t-SNE, Analytic Hierarchy Process, financial crime, FATF, text analytics, expert-in-the-loop learning

The framework, called an event-driven graph reinforcement learning system, tackles two long-standing weaknesses in AI-driven portfolio management. First, most deep reinforcement learning traders treat stocks as vectors of price indicators, only loosely connected or entirely independent. Second, when they do use news, they usually compress entire articles into a single sentiment score, positive, neutral or negative. That compression destroys exactly the information that matters during a crisis: who did what to whom, which company acquired which, which supplier is delayed, which bank was fined. A single polarity number cannot distinguish an acquisition from a regulatory action against the same firm, yet the two events demand opposite trading responses.

The researchers’ answer is a three-stage natural language processing pipeline built on financial language models. Raw headlines first pass through a FinBERT-based tone model that produces a magnitude-aware sentiment score, then through a zero-shot classifier that sorts each headline into one of eight event categories, from mergers and acquisitions to supply-chain disruptions, earnings reports, legal actions and macroeconomic news. Finally, a machine reading comprehension model, fine-tuned on a question-answering format, interrogates each headline with template questions such as “Which company is the acquirer?” or “Which company was fined?” The output is a structured event tuple capturing the primary entity, the counterparty, the event type, the sentiment magnitude and the model’s confidence in its own extraction.

Here is where the architecture departs from convention. Instead of appending these tuples as flat features, the system turns them into temporary nodes inside a heterogeneous market graph of 41 nodes: 28 stocks, 7 sector aggregates, one global macro node and a buffer of transient event nodes. Each trading day, the dynamic adjacency matrix is a weighted blend of a static, correlation-based topology and an event-driven topology built from that day’s extracted news. A merger headline creates a directed edge between two companies; a supply-chain story strengthens the link between a supplier and a manufacturer; a regulatory fine updates a bank’s risk features. Events decay exponentially with a half-life of roughly seven trading days, so stale news fades rather than contaminating later decisions. To avoid look-ahead bias, headlines published after 4 p.m. are deferred to the next day’s state, preserving temporal causality in the backtests.

On top of this living graph sits a Soft Actor-Critic agent, a reinforcement learning algorithm well suited to continuous control problems like allocating capital across 28 assets. The agent’s encoder uses two layers of a Hierarchical Graph Transformer with type-specific attention projections, so messages between stocks, sectors, macro nodes and events are processed differently depending on what they are. A hierarchical softmax policy head decomposes the allocation into a sector-level decision followed by a within-sector decision, guaranteeing by construction a fully invested, long-only portfolio. The reward function blends log wealth with penalties for oscillating weights and deep drawdowns, plus a small diversity bonus.

The risk machinery is the study’s most intricate contribution. The critic, the network that estimates the value of each action, is a dual-head Implicit Quantile Network, a distributional critic that models the full probability distribution of returns rather than just their mean. One head pools information exclusively from equity nodes while the other pools from macro, sector and event nodes; both are trained on the same reward, but their structural separation prevents risk-seeking equity gradients from overwriting defensive representations learned from macroeconomic signals. A curriculum on Conditional Value-at-Risk gradually amplifies the loss on the lowest quartile of the return distribution, so the agent first learns to seek returns and only later is forced to internalize tail risk. Two further filters, volatility-aware attention gating and momentum masking that blocks message passing from free-falling sectors, keep noise and contagion from flooding the graph.

The ablation experiments reveal how interdependent these components are. Adding the event graph and FinBERT pipeline to a static graph agent lifted the Sharpe ratio from 0.71 to 0.95 and more than doubled returns on the test window. But the risk components fail in isolation: the IQN critic alone actually degraded performance, because quantile sampling injects high-variance gradients, and a standalone attention gate suppressed useful signals. Only the integrated configuration controlled drawdowns effectively. A comparison against a simple keyword-based sentiment encoder was even starker: the FinBERT-driven model ended training with a Sharpe ratio 1.7 times higher and over 1,000 basis points less maximum drawdown, because binary keyword triggers simply cannot carry the magnitude and context a risk policy needs.

The headline results come from a standardized 90-day benchmark on a 28-stock universe spanning seven sectors. The event-driven agent achieved a Sharpe ratio of 2.11 and a Sortino ratio of 3.59, outperforming buy-and-hold, mean-variance, risk parity, an LSTM-SAC baseline and a replicated GraphSAGE-PPO model across all risk-adjusted metrics, with a modest 2.3 percent daily turnover. Notably, the replicated GraphSAGE-PPO baseline scored well below its originally published figures, a gap the authors attribute to differences in transaction cost assumptions and evaluation splits, a cautionary tale for the entire field of deep learning finance.

The most dramatic evidence comes from out-of-sample crisis stress tests on periods the model never saw. During the 2020 COVID-19 crash, the agent’s maximum drawdown was 26.23 percent, against roughly 33 percent for passive benchmarks, and its worst single-day loss was 9.98 percent versus about 11 percent for the baselines, all under a realistic cost model that charges a 25-basis-point fee plus VIX-scaled quadratic slippage that grows precisely when liquidity evaporates. The agent held a defensive core of telecom, staples, pharmaceutical and industrial names, and made one strikingly specific move: it bought Bank of America at a 1.6 percent allocation exactly at the crash trough on March 13, 2020, and exited as recovery began. Similar, if less dramatic, containment appeared in the 2015 China crash and the 2018 Volpocalypse. On a larger 50-asset point-in-time slice of the S&P 500, the event-driven model reached a Sharpe ratio of 2.55, while removing the event channel collapsed it to negative territory, a gap of 2.63 Sharpe units attributable to news alone.

Perhaps the most intriguing finding comes from the explainability analysis. A perturbation-based explainer showed that the decision to buy Bank of America was driven not by the bank’s own features, which ranked near the bottom in importance, but by signals propagated from defensive incumbents like Johnson & Johnson, Walmart and the industrial sector aggregate. The agent, in other words, was reading the graph, not the stock. The authors are candid about limitations: the framework has an inductive bias toward news-covered universes and underperformed a static-graph variant on nine news-starved assets, the slippage model remains a heuristic, and upstream extraction errors could in principle trigger unwarranted trades, though simulated corruption of up to 30 percent of headlines left the Sharpe ratio above 2.0. The team points toward temporal graph networks, multi-agent architectures and even physics-informed constraints as next steps. For now, the study offers a compelling demonstration that when the market’s structure changes, an AI trader should change its map with it, and that surviving a crash may depend less on predicting prices than on understanding the story spreading through the network.

Subject of Research: Event-driven graph reinforcement learning for risk-aware portfolio optimization

Article Title: Event-driven graph reinforcement learning for risk-aware portfolio optimization

Article References: B., P., & J., V. (2026). Event-driven graph reinforcement learning for risk-aware portfolio optimization. Machine Learning with Applications, 26, Article 101018. https://doi.org/10.1016/j.mlwa.2026.101018

Image Credits: AI Generated

DOI: 10.1016/j.mlwa.2026.101018

Keywords: reinforcement learning, graph neural networks, portfolio optimization, FinBERT, natural language processing, Conditional Value-at-Risk, Soft Actor-Critic, Implicit Quantile Network, market crashes, financial machine learning, sentiment analysis, risk management