This survey explores how explainable AI (XAI) techniques—such as SHAP, LIME, and interpretable models like attention-based LSTMs and decision trees—are applied to financial time series forecasting tasks like stock price prediction, volatility analysis, and algorithmic trading. By enhancing transparency and trust, these XAI approaches help financial professionals better understand model decisions, align predictions with market behavior, and support more informed investment strategies

This paper introduces a new type of regime-switching model that uses nonlinear kernel-based representations to automatically detect and forecast shifts in financial market behavior across multiple time series. Its key innovation lies in uncovering hidden, time-varying patterns without prior knowledge of regimes—making it highly relevant for financial applications like volatility forecasting, risk assessment, and market trend analysis.

This paper introduces a new type of regime-switching model that uses nonlinear kernel-based representations to automatically detect and forecast shifts in financial market behavior across multiple time series. Its key innovation lies in uncovering hidden, time-varying patterns without prior knowledge of regimes—making it highly relevant for financial applications like volatility forecasting, risk assessment, and market trend analysis.

This paper introduces a kernel-based self-representation learning method for discovering dynamic regimes in co-evolving time series, enabling more accurate and interpretable multi-step forecasting. In finance, this approach is particularly relevant as it captures nonlinear interactions and regime shifts—such as market volatility or economic cycles—without requiring prior knowledge, enhancing predictive modeling for complex financial systems.

This thesis proposes a correction method and a Bayesian estimator for the Pickands dependence function, improving accuracy and ensuring mathematical validity in modeling extreme events. Applications in finance include modeling joint extreme events such as simultaneous market crashes or co-movements in asset returns, improving risk assessment in portfolio management, and stress testing by accurately capturing tail dependencies between financial instruments. (In French)

This paper presents a framework using a temporal bipartite graph to model dependencies in time series. It uses autoregressive models for pattern identification and graph-based learning to capture transitions. The method aims to detect causal relationships in financial time series, comparing its accuracy and execution time with traditional methods like Granger causality and PCMCI. Designed for scalability and interpretability, its potential applications include risk assessment, portfolio management, and market behavior analysis in finance.

The paper uses AI to analyze thousands of YouTube videos from Bloomberg and Yahoo Finance, extracting financial insights through speech-to-text transcription and natural language processing. This helps identify key market trends, influential entities, and evolving narratives—offering valuable tools for financial analysis and investment decision-making.

This paper presents a new method for forecasting financial markets by identifying different market “regimes” using clustering across multiple time series. It captures nonlinear relationships between assets and adapts to changing conditions, leading to more accurate predictions. This approach helps investors and analysts understand shifts in market behavior, improving portfolio management, risk assessment, and trading strategies.

This paper proposes a novel model for financial market prediction that dynamically identifies cross-sectional regimes in multi-time-series data, allowing for the discovery of new market regimes as they emerge, rather than relying on a fixed set of pre-identified regimes. This is highly relevant in finance because it enhances the ability to detect structural market changes—such as those during crises or bubbles—improving prediction accuracy and offering better insights into market behavior through time-varying transition probabilities.

This paper introduces N-BEATS(P), a scalable and memory-efficient deep learning architecture for univariate time series forecasting that significantly reduces computational costs while maintaining state-of-the-art accuracy. Its relevance in finance lies in its demonstrated ability to perform zero-shot forecasting on financial datasets—such as stocks and ETFs—enabling rapid, cost-effective deployment of predictive models across diverse financial instruments without retraining.

This paper proposes STANN, a spatiotemporal adaptive neural network that improves long-term forecasting of multivariate financial time series by dynamically adjusting its autoregressive order using an attention mechanism. Applied to investment fund data, STANN significantly enhances forecasting accuracy and supports more effective autonomous trading strategies in finance.

This paper introduces an unsupervised deep generative model that improves financial time series forecasting by learning latent representations and inter-series relationships using a dynamic attention mechanism. This approach significantly enhances the accuracy of multi-asset forecasts, making it valuable for financial applications like ETF and mutual fund trajectory prediction.

This paper presents t a variable-order regime switching framework that identifies and predicts financial market regimes by extracting statistically significant behavioral patterns from time series data. Applied to volatility forecasting of 200 S&P 500 stocks, the model outperforms traditional regime-switching methods by offering interpretable insights into market dynamics and improving predictive accuracy.