Smart Forecasting: Deep Learning and Explainable AI Reshape Time Series Prediction

Artificial intelligence is quietly rewriting one of the oldest problems in data science: predicting what happens next. A new editorial published in Frontiers in Big Data on 10 September 2026, led by Leonard Barolli of the Fukuoka Institute of Technology together with Antonino Ferraro, Antonio Galli, and Francesco Moscato of universities in Naples and Salerno, argues that forecasting has entered a fundamentally new stage. The editorial introduces a Research Topic titled Smart forecasting: deep learning and explainable AI for real-world time series prediction, which gathers six studies spanning healthcare, software engineering, banking, and financial markets. What unites them is a striking shift in priorities: raw predictive accuracy, the editors contend, is no longer the finish line. Instead, practical forecasting systems must now be interpretable, adaptive, and genuinely useful to the human decision-makers who depend on them.

For decades, forecasting relied on statistical and econometric methods such as ARIMA, exponential smoothing, and GARCH-family models. These techniques remain powerful, but they struggle with the tangled temporal dynamics of modern data streams. The editorial traces how machine learning, deep learning, explainable artificial intelligence (XAI), and most recently large language models (LLMs) have expanded what predictive systems can do. Modern models no longer consume only structured numerical time series. They increasingly fuse heterogeneous information sources: textual reports, graph representations of interconnected entities, contextual knowledge, and domain expertise. This multimodal richness, fueled by growing data availability and advances in computational intelligence, allows forecasting models to build far richer representations of complex phenomena than classical statistics ever could.

Among the most eye-catching contributions is HairSentinel, a time-aware anomaly detection framework proposed by Leema and colleagues for forecasting hairfall trends. Rather than relying on image-based diagnosis, the system models longitudinal behavioral and physiological information collected through user interactions, using Temporal Fusion Transformers to identify abnormal hairfall patterns early. The approach is computationally efficient and accessible, illustrating how deep learning can support preventive healthcare by combining temporal modeling with anomaly detection. It is a vivid example of forecasting expanding beyond its traditional home in economics and finance into domains where anticipating future events can significantly improve early intervention and resource allocation.

Software engineering provides a second unexpected testing ground. Aversano and colleagues investigated forecasting bug resolution times in open-source software projects, comparing classical machine learning and deep learning approaches under both local and global forecasting scenarios across multiple repositories. Crucially, the study integrates explainable AI techniques, including permutation importance, saliency maps, and embedding analysis, to identify which software process characteristics most strongly influence prediction outcomes. The editors suggest that this combination of deep learning with explainability enables more transparent and actionable predictions in domains where operational planning and early intervention play a central role, offering project managers a data-driven window into how long defects might linger.

Finance, of course, remains the most dynamic arena for AI-driven forecasting, and two studies from the Saudi financial system highlight the complementary strengths of econometric and deep learning approaches. Alnajjar and colleagues examined the determinants of banking stability through an integrated framework combining econometric analysis with ARIMA and exponential smoothing models. Their findings indicate that capitalization, profitability, asset quality, and bank size significantly influence financial stability, while long-term forecasting provides useful support for banking supervision and policy planning within the Saudi Vision 2030 framework. Meanwhile, AL-Besher and AL-Najjar compared traditional GARCH-family models with Long Short-Term Memory (LSTM) neural networks for forecasting returns and volatility in the Saudi stock market. Their results show that GARCH models remain particularly effective for return prediction, whereas LSTM networks better capture nonlinear volatility dynamics, underscoring that neither paradigm should be discarded.

The emergence of large language models is opening perhaps the most transformative research direction. Wang and Yeung proposed an innovative LLM-driven forecasting framework for stock information networks that combines financial time series, institutional fund disclosures, retrieval-augmented generation, and graph representations. By jointly reasoning over structured numerical indicators and unstructured textual information, the methodology significantly improves forecasting accuracy while providing economically interpretable explanations of evolving market dynamics. The editors view this as evidence of a broader transition in which language models are evolving from natural language processing tools into general reasoning systems capable of supporting complex predictive tasks, particularly promising for forecasting applications characterized by dynamic, multimodal, and interconnected data.

Consolidating this rapidly moving field, Rohan and colleagues presented a comprehensive review of artificial intelligence for financial market prediction, tracing the evolution from traditional statistical approaches toward machine learning, deep learning, reinforcement learning, blockchain-enabled financial systems, and quantum computing. Beyond summarizing methodologies, the review examines challenges related to explainability, ethical AI, data quality, regulatory compliance, and hybrid intelligence, while outlining future directions for intelligent financial decision support. The editors position this review as a conceptual framework connecting the methodological advances of the empirical studies, highlighting how forecasting research is converging toward integrated AI ecosystems where predictive performance, transparency, adaptability, and human-centered decision support become equally important objectives.

Several common research trends emerge across the six contributions. First, forecasting is progressively evolving from purely numerical prediction toward intelligent decision-support systems capable of integrating structured time series, textual data, graph representations, and domain knowledge. Second, explainability is increasingly becoming a fundamental design requirement rather than an optional feature: several studies explicitly incorporate XAI techniques to improve model transparency, facilitate interpretation, and increase users’ trust in AI-assisted predictions. As forecasting systems become more deeply embedded in operational workflows, explainability will likely play a decisive role in their widespread adoption, particularly in scenarios where forecasting outcomes directly influence operational or strategic decisions.

Third, the collected studies consistently highlight the value of hybrid intelligence. Rather than viewing statistical methods and AI-based approaches as competing paradigms, they demonstrate that combining econometric models, machine learning, deep learning, graph analytics, and large language models often leads to more robust, accurate, and interpretable forecasting systems. The rise of foundation models suggests that future forecasting frameworks will increasingly combine reasoning capabilities with predictive modeling, enabling AI systems to process heterogeneous information while generating more meaningful explanations for human decision-makers. The editors argue that AI is progressively extending existing methodologies rather than replacing them, enabling richer representations of temporal phenomena across a wide variety of real-world applications.

Overall, the editorial concludes that smart forecasting is entering a new stage characterized by increasingly intelligent, explainable, and application-oriented predictive systems. Across healthcare, software engineering, banking, and financial markets, the studies consistently show that forecasting performance alone is no longer sufficient; practical adoption increasingly depends on the ability of AI systems to provide transparent, robust, and actionable predictions that effectively support human decision-making. The editors hope the Research Topic will stimulate further interdisciplinary research on trustworthy AI-driven forecasting systems that combine predictive accuracy with explainability, adaptability, and human-centered design. As forecasting continues to evolve alongside advances in artificial intelligence, these characteristics will likely become fundamental for the next generation of intelligent predictive systems.

Subject of Research: Editorial: Smart forecasting: deep learning and explainable AI for real-world time series prediction

Article Title: Editorial: Smart forecasting: deep learning and explainable AI for real-world time series prediction

Article References: Barolli, L., Ferraro, A., Galli, A., & Moscato, F. (2026). Editorial: Smart forecasting: deep learning and explainable AI for real-world time series prediction. Frontiers in Big Data, 9, Article 1954626. https://doi.org/10.3389/fdata.2026.1954626

Image Credits: AI Generated

DOI: 10.3389/fdata.2026.1954626

Keywords: Editorial, Smart, forecasting, deep, learning, explainable, real-world, time, series, prediction, scientific research