Interpretable ESG-sentiment hybrid deep learning for asset return forecasting with quantified interactions and latency-aware deployment.

Skip to main content An official website of the United States government Here's how you know Here's how you know Official websites use .gov A .gov website belongs to an official government organization in the United States. Secure .gov websites use HTTPS A lock ( Lock Locked padlock icon ) or https:// means you've safely connected to the .gov website. Share sensitive information only on official, secure websites. Search Log in Dashboard Publications Account settings Log out Search… Search NCBI Primary site navigation Search Logged in as: Dashboard Publications Account settings Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide PERMALINK Copy As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice Sci Rep . 2026 Mar 4;16:12001. doi: 10.1038/s41598-026-41985-3 Search in PMC Search in PubMed View in NLM Catalog Add to search Interpretable ESG–sentiment hybrid deep learning for asset return forecasting with quantified interactions and latency-aware deployment Sasmita Mishra Sasmita Mishra 1 Department of Business Management, C. V. Raman Global University, Bhubaneswar, India Find articles by Sasmita Mishra 1 , Zefree Lazarus Mayaluri Zefree Lazarus Mayaluri 2 Department of Electrical Engineering, C. V. Raman Global University, Bhubaneswar, India Find articles by Zefree Lazarus Mayaluri 2, ✉ , Chee Yoong Liew Chee Yoong Liew 3 Department of Accounting and Finance, UCSI University, Kuala Lumpur, Malaysia Find articles by Chee Yoong Liew 3 , Prabodh Kumar Sahoo Prabodh Kumar Sahoo 4 Department of Mechatronics Engineering, Parul Institute of Technology, Parul University, Vadodara, India Find articles by Prabodh Kumar Sahoo 4, ✉ , Aswini Kumar Samantaray Aswini Kumar Samantaray 5 Department of Electronics and Communication Engineering, Manipal Institute of Technology Bengaluru, Manipal Academy of Higher Education, Manipal, India Find articles by Aswini Kumar Samantaray 5, ✉ Author information Article notes Copyright and License information 1 Department of Business Management, C. V. Raman Global University, Bhubaneswar, India 2 Department of Electrical Engineering, C. V. Raman Global University, Bhubaneswar, India 3 Department of Accounting and Finance, UCSI University, Kuala Lumpur, Malaysia 4 Department of Mechatronics Engineering, Parul Institute of Technology, Parul University, Vadodara, India 5 Department of Electronics and Communication Engineering, Manipal Institute of Technology Bengaluru, Manipal Academy of Higher Education, Manipal, India ✉ Corresponding author. Received 2025 Dec 11; Accepted 2026 Feb 24; Collection date 2026. © The Author(s) 2026 Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/ . PMC Copyright notice PMCID: PMC13068911  PMID: 41781523 Abstract Accurate forecasting of financial time series increasingly relies on alternative data such as environmental, social and governance (ESG) scores and news-based sentiment, yet the way these signals interact and when they actually improve forecasts is still poorly understood. We introduce an interpretable hybrid framework for asset return forecasting that combines a Temporal Fusion Transformer (TFT) with a lightweight Support Vector Regression (SVR) residual corrector and an explicit gated late fusion of ESG features with aspect-based financial sentiment (FinBERT-based ABSA). The gating mechanism learns when to emphasize sustainability versus sentiment signals, while SHAP interaction values and Friedman’s H quantify ESG–sentiment interactions across assets and regimes. A finance-grade, leak-proof walk-forward protocol (252 trading days train / 10 days test, within-fold scaling, ABSA items strictly before 16:00 ET; ESG effective T+3; macro T+1, HAC-robust Diebold–Mariano tests) is applied to US large-cap technology equities, major global indices, and BTC/ETH over 2020–2024. Across independent seeds, the hybrid achieves aggregate mean absolute error of and RMSE of on next-day log returns, with directional accuracy , IC 0.39, and ICIR 0.82, significantly outperforming tuned deep-learning and machine-learning baselines (HAC-robust per-asset Diebold–Mariano tests with BH-FDR ; Fisher aggregation yields ). Simple long-only, thresholded simulations indicate higher risk-adjusted performance and lower maximum drawdown under conservative transaction-cost assumptions. Ablation studies show that removing either ESG or sentiment features yields the largest degradations, and that the SVR corrector stabilizes errors under regime shifts. To directly address market-cycle sensitivity, we evaluate stability across event-defined stress windows (COVID-19 crash, 2022 tightening cycle, and 2023 banking stress) and volatility-defined regimes using terciles of 20-day realized volatility. We report regime-split forecasting and strategy metrics with block-bootstrap confidence intervals, HAC-robust Diebold–Mariano tests within each regime, and residual-stabilization diagnostics that quantify the SVR variance and skewness reduction under stress. ESG–sentiment interactions are statistically non-zero and regime-dependent, with sentiment gaining importance in turbulent periods and ESG in calmer markets. A latency-optimized variant that removes auxiliary BiLSTMs retains over of the accuracy gains while reducing inference time by approximately of the full model (i.e., a reduction of about ), supporting near-real-time deployment. Keywords: Financial time-series forecasting, ESG metrics, Aspect-based sentiment analysis, Temporal fusion transformer, Support vector regression, Explainable AI Subject terms: Engineering, Mathematics and computing Introduction Forecasting financial markets and constructing robust portfolios remain difficult because returns are volatile, nonstationary, and prone to regime shifts. Classical views—such as the Efficient Market Hypothesis and the Random Walk Hypothesis—limit the value of purely price–based signals 1 – 3 , motivating models that learn nonlinear structure and adapt to evolving conditions. Beyond traditional factors, two information streams have become prominent drivers of return dynamics and risk: (i) firm–level Environmental, Social, and Governance (ESG) metrics associated with resilience and long–horizon performance, and (ii) investor sentiment extracted from news and social media 4 – 6 . Episodes such as meme–stock rallies and ESG–related repricing show how sustainability narratives and collective psychology can rapidly propagate to prices—especially during stress. Because forecasting skill can vary substantially across market cycles, we explicitly test whether the proposed gains persist across distinct volatility regimes, including crisis and tightening episodes. We evaluate both event-defined stress windows and volatility-defined terciles and complement accuracy with statistical tests and diagnostic evidence (residual stabilization and interaction shifts) to ensure that improvements are not driven by a single market phase. Modern ML/DL architectures (ANNs, SVMs, LSTMs, and transformers) improve on linear baselines, yet when deployed in isolation they often struggle to reconcile heterogeneous signal types whose relevance is regime–dependent 7 – 11 . In particular, prior hybrids frequently include ESG or sentiment but rarely model their interaction explicitly, and evaluation protocols sometimes omit finance–specific safeguards (strict walk–forward splits, leakage controls, Diebold–Mariano tests, and strategy–relevant metrics). Contributions. We address these gaps with a compact and interpretable framework that: introduces an explicit, gated late–fusion between ESG and aspect–based sentiment (ABSA), enabling regime–aware reweighting of sustainability versus mood signals; interaction is quantified via SHAP interaction values and Friedman’s ; employs a finance–grade, leak–proof walk–forward evaluation (252/10 rolling windows; within–fold scaling; as-of lags (ABSA items strictly before 16:00 ET; ESG effective T+3; macro T+1); Diebold–Mariano tests) and reports both statistical and economic value (Sharpe, Sortino, MDD); provides a practitioner–ready variant (TFT + SVR; no BiLSTM) that preserves of the accuracy benefits while reducing latency by , improving deployability without sacrificing interpretability. evaluates stability across market cycles using event-defined stress windows (COVID-19 crash, 2022 tightening, 2023 banking stress) and volatility-defined regimes (20-day realized-volatility terciles), reporting regime-split accuracy and economic metrics together with HAC-robust Diebold–Mariano tests; and provides mechanism-level regime diagnostics via gate dynamics and regime-sliced SHAP interaction change tests, and quantifies SVR residual stabilization (variance and skewness reduction) by regime. Overview. We integrate technical indicators, macro variables, ESG scores, and FinBERT–ABSA sentiment within a Temporal Fusion Transformer (TFT), and use a lightweight SVR to correct residuals. We benchmark against recent alternatives—including DeepTVAR and the LLM–based FinGPT—on US technology equities, major global indices, and BTC/ETH (2020–2024). Although average point–error gains over FinGPT are modest, we observe material risk improvements (lower MDD; higher Sharpe/Sortino) and reduced error dispersion across rolling windows. Interaction evidence indicates that the ESG–sentiment Friedman H statistic has a median of across assets (IQR 0.18–0.24), and ESG–sentiment SHAP interaction values are consistently non-zero across folds, peaking during high-volatility windows. Literature review Traditional financial forecasting models Classical econometric models (linear regression, ARIMA, GARCH) remain valuable for transparency and well-understood assumptions 7 , 12 . However, performance degrades under nonstationarity, structural breaks, and volatility spikes (e.g., COVID-19), where model misspecification and lagged adaptation are common 13 – 15 . Their limited capacity to capture nonlinear, regime-dependent dynamics has motivated data-driven alternatives. Machine learning and deep learning Machine-learning and deep-learning approaches (ANNs, SVMs) and sequence models (LSTM/BiLSTM, transformers) improve nonlinear fit and temporal-dependence modeling 8 , 16 – 21 . Yet single-family models often struggle when the relevance of signals shifts across regimes, producing brittle behavior during turbulence 22 , 23 . This has spurred hybrids that combine denoising, sequence learning, and residual correction. Hybrids: strengths and gaps Hybrid paradigms—e.g., EEMD–LSTM–SVR, DeepAR, and transformer-based text models (BERT/FinBERT)—blend complementary inductive biases 5 , 11 , 22 , 24 – 28 . N-BEATS and attention-based ensembles add scalability and flexible pattern extraction 29 . Recent hybrid pipelines increasingly combine heterogeneous feature extraction with automated tuning. Convolutional backbones (e.g., CNN/TCN) can extract local patterns from multichannel market representations, recurrent units (e.g., GRU/LSTM) capture temporal dependencies, and global optimizers (e.g., genetic algorithms) search hyperparameters that are sensitive under manual/grid tuning. Ke et al. present a GRU–CNN–GA fusion exemplar with GA-driven hyperparameter optimization in a hybrid forecasting pipeline 30 . Bayesian optimization frameworks are also used to reduce manual tuning overhead in recent deep forecasting systems 31 . Three persistent gaps remain: Interaction blindness. ESG and sentiment are frequently appended as separate features; a few studies use multi-modal attention to implicitly fuse text with market variables (e.g., 5 , 10 ), but explicit interaction modeling and quantification (e.g., SHAP interactions, Friedman’s ) remain rare, limiting economic interpretation and regime-aware reweighting. Evaluation rigor. Strict walk-forward validation, leakage controls, Naïve persistence comparators, and formal significance tests (Diebold–Mariano) are inconsistently applied; portfolio metrics are sometimes misused for point-error comparisons 32 – 35 . Deployability. State-of-the-art stacks can be compute-heavy and brittle to tune; simplified, practitioner-ready variants are under-documented 36 . A concise comparison of representative hybrids—covering strengths, limitations, and suggested improvements—is summarized in Table 1 . Table 1. Comparison of leading hybrid approaches. Approach Strengths Limitations Typical use cases Suggested improvements EEMD–LSTM–SVR Combines noise reduction, sequential learning, and robust regression Higher architectural complexity and computational overhead High-noise markets (e.g., cryptocurrencies) Dynamic feature selection; SHAP-based interpretability DeepAR DL scalability with autoregressive structure Weaker adaptability to structural breaks and high-frequency volatility Retail sales; macroeconomic forecasting Attention-aided volatility modeling Transformer (BERT/FinBERT) Strong contextual modeling for text/sentiment Data-hungry; limited transparency Sentiment-augmented forecasting; financial NLP Distilled/lightweight variants; XAI integration N-BEATS Scalable and generic; effective on multivariate series Limited domain-specific interpretability Broad, cross-sector forecasting Domain adaptation; interpretability layers Attention-based LSTM ensembles Dynamic reweighting; resilience under volatility Overfitting risk in small samples; opaque attributions Sector-level stock forecasting Stronger regularization; post-hoc explainability GRU–CNN–GA fusion CNN captures local patterns while GRU models temporal dependence; GA-based search reduces manual tuning sensitivity 30 Higher training cost; often tailored to high-frequency or multi-stream inputs; limited transparency without XAI Heterogeneous feature extraction; automated hyperparameter optimization in hybrid forecasting Add interaction analysis for heterogeneous signals; leak-proof walk-forward evaluation; latency-aware variants Open in a new tab Sentiment, ESG, and alternative data Transformer-based sentiment (FinBERT) enables finance-specific ABSA features 10 , 27 , 37 , while ESG has been linked to risk resilience and long-horizon quality 24 , 38 . Most studies, however, integrate either ESG or sentiment—or fuse them without an adaptive mechanism. Asset coverage is often narrow, and protocols for mixed-frequency alignment (quarterly ESG vs. daily prices) and lagging of text signals are not always explicit 39 , 40 . Positioning of the present work Two common fusion strategies appear in the literature. (i) Feature concatenation , where ESG and sentiment are appended to technical/macroeconomic inputs (typical in FinBERT-based forecasting streams 26 , 27 and in hybrid modeling pipelines 22 , 24 ), treats ESG and sentiment as additive covariates but does not model their joint effect or its regime dependence. (ii) Multi-modal attention , which can implicitly reweight modalities (e.g., transformer-based financial NLP and robustness-focused variants 5 , 10 ), improves flexibility but usually stops short of quantifying cross-modal interaction or enforcing finance-grade evaluation (strict walk-forward, within-fold scaling, explicit as-of lags). By contrast, our gated late-fusion (a) places an explicit scalar gate on the ESG–sentiment channel to capture their interaction and regime-aware trade-off, (b) quantifies that interaction ex post via SHAP interaction values and Friedman’s , and (c) evaluates under a leak-proof, finance-specific protocol (252/10 rolling splits; within-fold scaling; as-of lags (ABSA items strictly before 16:00 ET; ESG effective T+3; macro T+1); HAC-robust Diebold–Mariano tests). Empirically, this design reveals a statistically non-zero ESG–sentiment interaction (median ) and clarifies when sentiment (turbulent windows) versus ESG (stable windows) dominates, turning fusion into a testable economic statement rather than a black-box aggregation. Methodology Workflow overview We develop a compact hybrid pipeline that integrates (i) multi-modal temporal learning via a Temporal Fusion Transformer (TFT), (ii) a lightweight Support Vector Regression (SVR) residual corrector, and (iii) an explicit, gated late-fusion of Environmental, Social, and Governance (ESG) features with aspect-based sentiment (ABSA). The design targets three objectives: (a) accuracy under nonlinear dynamics and regime shifts, (b) explicit modeling and quantification of ESG–sentiment interactions, and (c) deployability enforced by clear leakage guards. A schematic is provided in Fig. 1 . Figure 1. Open in a new tab Schematic overview of the proposed hybrid financial forecasting and portfolio optimization framework. The pipeline integrates multi-source data (technical, ESG, sentiment), advanced deep learning and machine learning modules, and model explainability analysis. Data collection and preprocessing We study US large-cap tech equities (AAPL, AMZN, MSFT, NFLX), major global indices, and BTC/ETH. All series are daily and aligned to a unified as-of 16:00 ET framework to avoid look-ahead. Key sources and lags appear in Table 2 . Table 2. As-of alignment (all series at 16:00 ET). Domain As-of rule/lag Equities/indices 16:00 ET close; NYSE calendar; tech features from train-only windows Crypto (BTC, ETH) Aggregate 00:00–23:59 UTC; shift to 16:00 ET (same civil date) Sentiment (ABSA) Items strictly before 16:00 ET; weekends/holidays roll; lag: same day ESG Publication stamp + T+2 embargo, then +1 trading day (effective T+3) Macro (monthly) Forward-filled to daily within fold; available T+1 Open in a new tab Time harmonization. Timestamps are converted to America/New_York. Equities/indices use NYSE regular-session close (16:00 ET); Binance crypto is aggregated over 00:00–23:59 UTC and shifted to align with 16:00 ET on the same civil date. News/ABSA features include only items strictly before 16:00 ET (weekends/holidays roll to the next trading day). ESG follows vendor publication stamps with a conservative embargo (see Table 2 ). Leakage guards (per rolling fold). Walk-forward: 252 train / 10 test; window advances by 10 days. Scaling: z-scores fit on train-only data; applied to test. As-of lags: ABSA same-day pre-close; ESG effective T+3; macro T+1. Features: technical indicators computed within fold from train windows only. Naïve persistence is the baseline 41 . Missingness 2% is imputed within each fold using past-only forward filling (no future values are used); imputation is never performed across market closures. Regime definitions (for stability analysis). To assess robustness to market cycles, we define (i) event windows : COVID-19 crash (2020-02-20 to 2020-04-30), 2022 tightening cycle (2022-01-01 to 2022-12-31), and 2023 banking stress (2023-03-01 to 2023-05-31); and (ii) volatility regimes using terciles of 20-day realized volatility computed from daily log returns. Volatility terciles (low/mid/high) are assigned within each walk-forward fold to avoid look-ahead; these regimes are used only for slicing evaluation and do not alter training splits. We use the full calendar year 2022 as a conservative tightening-cycle window to avoid post-hoc cherry-picking of subperiods; finer sub-windowing is left for future work. Feature engineering We construct three principal feature domains. Technical indicators. MACD, RSI, Bollinger Bands, EMA/SMA, ATR, etc., computed with standard formulas; all technical features are z-scored per fold. PCA is fit on the train portion of each fold only and applied to that fold’s test window ; PCA is used only for technicals (ESG/ABSA kept in native form for interpretability). ESG variables. Aggregate E/S/G scores (Bloomberg) undergo IQR-based outlier filtering ( ) and linear interpolation where needed, then are z-scored per fold. Values are aligned & lagged one trading day (effective T+3 availability; see Data section). ABSA sentiment (FinBERT). We fine-tune FinBERT on FiQA and compute volume-aware daily asset scores to avoid news-volume bias. Concretely, for each asset/day we: (i) collect items timestamped strictly before 16:00 ET, (ii) score sentences, then aggregate with a content-weighted mean (doc length as proxy), and (iii) When , the prior day’s ABSA value is carried forward and a missingness flag is retained as an additional feature. The FinBERT fine-tuning configuration used to obtain aspect-level sentiment scores is summarized in Table 3 . Table 4 reports aspect metrics. Table 3. FinBERT fine-tuning (concise). Dataset / task FiQA Aspect Polarity (v2); 80/10/10 (asset-stratified) Labels {neg, neu, pos}; class weights by inverse frequency Model / optimizer FinBERT-base; AdamW (weight decay 0.01) LR / batch / epochs / 16 / 5 (early stop patience 2) Max seq length 128; truncation with whole-word mask Seed / hardware 42 / single GPU (VRAM 12 GB) Selection metric Macro-F1 on validation Open in a new tab Seed fixed for reproducibility. Table 4. FinBERT aspect performance on FiQA (validation). Aspect Acc. (%) F1 AUC Company perf. 95.2 0.93 0.95 Macro sentiment 93.7 0.91 0.92 Market trends 94.4 0.92 0.94 Industry news 94.0 0.91 0.93 Other topics 93.5 0.90 0.92 Overall 94.1 0.92 0.94 Open in a new tab Dimension control. All technical indicators are z-scored using train-only statistics within each fold. PCA is then fit on the training portion of each fold, and the resulting transform applied to the corresponding test window. The number of retained components is selected per fold on the training window using the Bai–Ng criterion 42 , 43 , ensuring fold-specific and leak-proof dimension control. ESG and ABSA features are retained in native form for interpretability. ESG–sentiment fusion and interaction quantification Let , , and . We form ESG and sentiment embeddings 1 with . A scalar gate combines them: 2 and the fused input is 3 which is fed to the temporal module with a 30-day lookback. We use a scalar gate for parsimony and training stability; a vector gate (elementwise ) yields similar accuracy but higher variance/compute. ESG–sentiment interaction is quantified ex post via SHAP interaction values and Friedman’s H ; model-agnostic ALE plots provide complementary effect shapes. Model architecture and integration Our deployable default is a two-stage architecture: Temporal module (TFT). Two transformer layers, four attention heads, dropout , 30-day lookback; input is . Residual corrector (SVR). An RBF-kernel SVR consumes to correct predictable residuals under regime shifts and nonlinearities. An auxiliary BiLSTM (two layers, 128 units) is examined in ablations; we recommend TFT + SVR in practice due to substantially lower latency while retaining of the full-model accuracy. Hyperparameter tuning and optimization TFT. Manual/grid search within prior-informed ranges: layers , heads , dropout ; early stopping on fold-level validation (patience epochs; min ). SVR. Bayesian optimization per fold (budget trials; warm start random) using a Tree-structured Parzen Estimator (TPE) surrogate with Expected Improvement acquisition. Priors are log-uniform: , , . Median selected values across assets/folds were , , ). Figure 2 shows convergence: TPE outperforms random and coarse grid on mean error reduction per budget. Figure 2. Open in a new tab SVR HPO Convergence. Mean best-so-far normalized validation loss vs. evaluation budget (50 per method). Shaded bands show 95% bootstrap CIs across runs (assets/folds/seeds). Bayesian (TPE+EI) reduces loss fastest; Random and Grid converge more slowly under the same budget. ABSA fine-tuning. Learning rate , batch size , early stopping (max 5 epochs), AdamW ( ), seed . The tuned hyperparameter ranges and the chosen (median) settings for the deployable TFT+SVR configuration are reported in Table 5 . Table 5. Hyperparameters for the deployable TFT+SVR (BiLSTM appears in ablations). Component Parameter Range/method Chosen/median TFT Layers, heads ; (grid) 2; 4 Dropout (grid) 0.2 SVR (RBF) TPE+EI; log-uniform priors; 50 trials/fold FinBERT LR, batch, epochs ; ; early stop Open in a new tab Evaluation and statistical testing We forecast next-day log returns (horizon ) on a leak-proof 252/10 walk-forward with fold-wise scaling and as-of lags (ABSA items strictly before 16:00 ET; ESG effective T+3; macro T+1). Primary point metrics are MAE and RMSE on returns; directional accuracy (DA) is computed directly on the sign of log returns. For visualization, price paths can be reconstructed as . 4 5 6 Multi-seed protocol and confidence intervals. All headline metrics are averaged over independent runs with different random seeds for data shuffles and model initialization. We report mean ± 95% confidence intervals (CI) obtained via non-parametric bootstrap over folds (and seeds where applicable). Error bars in figures denote mean ± 95% CI unless stated otherwise. Statistical significance of model-vs-baseline loss differentials uses HAC-robust Diebold–Mariano tests with Newey–West bandwidth (standard rule-of-thumb; is the test-window length per comparison). We adjust per-asset -values across multiple models via Benjamini–Hochberg FDR at . As a sensitivity check, we aggregate DM statistics across folds using a stationary block bootstrap (block length trading days). Regime-split evaluation and stability tests. In addition to full-sample metrics, we compute MAE, RMSE, DA, IC, and ICIR within each event-defined and volatility-defined regime. Economic metrics (Sharpe, Sortino, maximum drawdown, turnover) are computed using the same long-only threshold rule described in “ Economic value ”, applied within each regime. Statistical comparisons between TFT+SVR and FinGPT are repeated within each regime using HAC-robust Diebold–Mariano tests on squared-error loss; -values can be reported raw and optionally adjusted across assets via BH-FDR ( ). To quantify regime stability, we report the difference between high- and low-volatility MAE (and RMSE) and assess significance via paired block bootstrap over folds (block length 5 trading days). Operating-point metrics, calibration, and paired tests. In addition to point errors (MAE, RMSE) and DA, we report: (i) precision–recall AUC (PR-AUC) for the sign task; (ii) sensitivity at fixed specificity thresholds (90% and 95%), with thresholds chosen on validation only; and (iii) probability calibration for the event . Continuous return forecasts are mapped to using Platt scaling (logistic regression) fit on validation data within each fold; calibrated probabilities are evaluated on the corresponding test window. Calibration is summarized with reliability diagrams, Expected Calibration Error (ECE), and Brier score, each with 95% CIs from paired block bootstrap over folds (block length 5 trading days). For paired comparisons on the sign task we use McNemar’s test with continuity correction; differences in PR-AUC and calibration metrics are assessed via paired block bootstrap and reported as 95% CIs. Explainability We compute SHAP attributions on the fitted TFT+SVR for technical, ESG, and ABSA inputs, and report SHAP interaction values for ESG–sentiment pairs together with Friedman’s H to quantify cross-modal interaction strength. For non-tree models we use KernelSHAP on the final-step inputs of each lookback window: the background set consists of 500 samples drawn from the training split of each fold, stratified by realized-volatility terciles; we repeat explanations over seeds and report mean ±95% CI. We assess faithfulness with deletion/insertion curves (probability-mass masking schedule) and summarize areas under the curves (AUC-d, AUC-i). Across assets, ESG–sentiment interactions are statistically non-zero (median ; IQR [0.18, 0.24]; permutation test with 1,000 label permutations per asset; BH-FDR ). Model-agnostic ALE plots provide complementary effect shapes. Figure 3 illustrates the time-varying attribution profile around key market events. Figure 3. Open in a new tab Temporal SHAP analysis showing changing feature importance for sentiment, ESG, macro, and technical indicators, with key market events annotated. Regime-conditioned mechanism analysis. To provide mechanism-level evidence for regime adaptivity, we examine (i) the dynamics of the learned fusion gate over time and (ii) regime-sliced changes in the absolute ESG ABSA interaction magnitude across realized-volatility terciles (low/mid/high) (Figs. 4 and 5 ). We test whether the distributions of and differ between low- and high-volatility regimes using a two-sided permutation test (1,000 permutations per asset) with BH-FDR control ( ). Figure 4. Open in a new tab Gate dynamics over time with stress-window shading. Monthly median gate value (points) with a 3-month rolling mean (line). Shaded regions denote event-defined stress windows (COVID-19 crash, 2022 tightening cycle, 2023 banking stress). Higher indicates greater relative emphasis on the ESG channel in the gated late-fusion, whereas lower indicates a shift toward sentiment. Figure 5. Open in a new tab Interaction change across volatility regimes (terciles of realized volatility). Distribution of the absolute ESG sentiment SHAP interaction magnitude across low/mid/high volatility terciles. The high-versus-low median shift is annotated as , alongside the two-sided permutation-test p -value (1,000 permutations; seed=42). Boxes show IQR with median; whiskers extend to 1.5 IQR. Experiments and results Setup and protocol We follow the leak-proof protocol in “ Evaluation and statistical testing ”: 252/10 walk-forward splits, fold-wise scaling (train-only), ABSA items strictly before 16:00 ET; ESG effective T+3; macro T+1, and train-only feature computation. Benchmarks include Naïve persistence, LSTM 44 , BiLSTM 45 , N-BEATS 29 , Informer 31 , DeepAR 24 , FinGPT 46 , and our deployable default TFT+SVR . (BiLSTM appears only in ablations.) The asset set spans US large-cap tech equities (AAPL, AMZN, MSFT, NFLX), global indices (S&P 500, NASDAQ Composite, FTSE 100, Nikkei 225, MSCI World), and BTC/ETH (Jan 2020–Dec 2024). We forecast next-day log returns; primary point metrics are MAE and RMSE on returns, complemented by DA, information coefficient (IC), and ICIR. Statistical significance is assessed via HAC-robust Diebold–Mariano tests on return-loss differentials (“ Evaluation and statistical testing ”). Economic value (Sharpe, Sortino, MDD, turnover) is computed from a transparent long-only, thresholded strategy on predicted returns. Reproducibility. We fix multiple random seeds across data splits and model initializations and average results over seeds. FinGPT baseline (text-only). We use FinGPT v3.3 (checkpoint: FinGPT-3.3-base) in few-shot mode with three exemplars per asset. Inputs consist of news headlines and short articles formatted as: “Asset: [Ticker]. Date: [YYYY-MM-DD]. News: [text]. Task: Predict next-day log return.” FinGPT outputs the next-day log return directly; prompt templates were selected via validation directional accuracy on 2020–2021 folds (approximately 20 prompt trials). FinGPT does not consume structured technical or ESG features and is therefore treated as a text-only auxiliary baseline. Only news items timestamped strictly before 16:00 ET are included. Few-shot exemplars are drawn from training-side periods preceding each test window, and prompt templates are selected using only 2020–2021 folds; no test-window text is used in prompt selection or calibration. Core environment: Python 3.10, PyTorch 2.x, scikit-learn 1.x; experiments run on a single NVIDIA-class GPU with CUDA 12.x. Costs. Strategy results are reported at zero explicit costs and stress-tested at 5 and 10 bps one-way transaction costs; qualitative model rankings and risk metrics remain stable under these levels. Headline accuracy and effect sizes As summarized in Table 6 , TFT+SVR improves both error metrics and DA relative to strong baselines (including FinGPT). Table 6. Aggregate return-forecasting performance across assets (2020–2024). Model MAE ( ) RMSE ( ) DA (%) IC (Pearson) ICIR DM p (vs Naïve) Naïve (Persistence) 72.4 0.05 0.12 Ref. LSTM 44 88.7 0.28 0.65 BiLSTM 45 89.2 0.30 0.68 N- BEATS 29 92.1 0.34 0.72 Informer 31 92.5 0.35 0.74 DeepAR 24 91.8 0.33 0.70 FinGPT 46 93.4 0.37 0.78 TFT+SVR (ours) 94.5 0.39 0.82 Ablation (no ESG) 90.2 0.27 0.61 Ablation (no Sentiment) 90.8 0.29 0.64 Open in a new tab Metrics computed on next-day log returns; values are mean ±95% CI across folds and seeds. Bold indicates best. DM -values compare each model to the Naïve (persistence) baseline using squared-error loss on returns (HAC/Newey–West). Directional diagnostics are summarized in Table 7 , confirming improved operating-point sensitivity. Table 7. Directional metrics (aggregate across assets; mean ±95% CI over folds and seeds). Model PR-AUC Sens@90%Spec Sens@95%Spec FinGPT TFT+SVR Open in a new tab Diebold–Mariano (DM) details. We compare daily squared-error losses on returns. For each asset, we compute HAC (Newey–West) DM statistics with bandwidth , where is the test-window length. Across the evaluated universe of 11 assets (4 US large-cap equities, 5 global indices, and 2 crypto), TFT+SVR significantly outperforms FinGPT in 9/11 cases under HAC-robust Diebold–Mariano tests after Benjamini–Hochberg FDR control ( ). Aggregated across assets, Fisher’s method yields a combined -value . DA computation details (clarification). DA is computed on all test days in each walk-forward fold (not only on days when the trading rule takes a position). Specifically, DA uses the sign of the next-day log return, i.e., versus , aggregated across all test observations, assets, folds, and seeds. Days with exactly zero realized return are rare in the studied daily series and are treated as correct only when both predicted and realized signs are zero. To avoid inflated interpretation from class imbalance, the fraction of positive-return days in the pooled test set is approximately balanced (reported in the reproducibility artifact), so DA should be interpreted jointly with MAE/RMSE, IC/ICIR, and HAC-robust DM tests rather than as a standalone metric. Figure 6 visualizes the fold-level distribution of MAE and RMSE across competing models, complementing the aggregate results reported in Table 6 . Figure 6. Open in a new tab Fold-level error distributions for MAE (left) and RMSE (right) across models. Boxes show interquartile range with median; whiskers extend to 1.5 IQR; diamonds mark the mean. Regime stability across market cycles Because forecasting performance is sensitive to market regimes, we evaluate stability across event-defined stress windows (COVID-19 crash, 2022 tightening, 2023 banking stress) and volatility-defined regimes using terciles of 20-day realized volatility. Table 8 reports regime-split forecasting metrics. TFT+SVR shows a statistically significant advantage over FinGPT at the regime level across all regimes (Fisher-combined HAC-DM p -values), with the largest absolute error reductions in high-volatility periods. Table 8. Regime-split forecasting performance. Regime Model MAE RMSE DA (%) IC ICIR HAC-DM p COVID(2020-02-20–04-30) TFT+SVR 91.5 0.34 0.68 0.012 FinGPT 87.2 0.21 0.42 – Naïve 81.0 0.05 0.10 – 2022 tightening (2022-01-01–12-31) TFT+SVR 93.5 0.38 0.76 0.008 FinGPT 90.2 0.27 0.54 – Naïve 85.0 0.10 0.20 – 2023 banking stress (2023-03-01–05-31) TFT+SVR 92.2 0.36 0.72 0.015 FinGPT 88.9 0.24 0.48 – Naïve 83.5 0.08 0.16 – Low vol tercile (20-day RV) TFT+SVR 95.5 0.42 0.84 0.004 FinGPT 92.0 0.30 0.60 – Naïve 88.0 0.12 0.24 – Mid vol tercile (20-day RV) TFT+SVR 93.0 0.35 0.70 0.009 FinGPT 89.5 0.25 0.50 – Naïve 84.0 0.09 0.18 – High vol tercile (20-day RV) TFT+SVR 90.1 0.31 0.62 0.010 FinGPT 86.0 0.22 0.44 – Naïve 79.0 0.04 0.08 – Open in a new tab Metrics are mean ±95% CI computed via block bootstrap across folds and seeds (block length = 5 trading days). Note: DA is computed on all test days (not only traded days) using the sign of next-day log returns. HAC-DM tests are computed per asset within each regime on squared-error loss with Newey–West bandwidth as in “ Evaluation and statistical testing ”; the table reports the Fisher-combined p -value across assets for each regime. BH-FDR ( ) is applied to per-asset DM p -values when reporting asset-level significance Economic value We evaluate a simple, transparent trading rule driven by predicted next-day returns. At time , take a long position if and otherwise hold cash, where is the 70th percentile of in-sample predicted returns (per asset, per fold). The threshold filters weak signals and reduces churn. Turnover is computed as the time-average of , where is the daily position indicator under the rule. Positions are held for 1 trading day with rolling rebalancing; position size is 1 notional (no leverage, no volatility scaling, no shorting). Equity curves are computed net of one-way costs bps. Sharpe and Sortino are annualized using ; Maximum Drawdown (MDD) is computed on the daily equity curve. Headline strategy metrics under the long-only threshold rule, aggregated across assets over 2020–2024, are reported in Table 9 . Table 9. Headline strategy metrics aggregated across assets (2020–2024) under the long-only threshold rule. Model Sharpe Sortino MDD (%) Turnover FinGPT 1.42 1.95 0.62 TFT+SVR 1.58 2.12 0.59 Open in a new tab Under this rule, TFT+SVR improves risk-adjusted performance and reduces drawdown relative to the text-only baseline, while maintaining comparable turnover. To assess whether the economic gains persist across market cycles, Table 10 reports the same strategy metrics split by event-defined and volatility-defined regimes. Table 10. Strategy metrics by regime for the long-only threshold rule (annualized Sharpe/Sortino; MDD reported as negative percent; turnover annualized), reported at 0 bps one-way transaction cost unless otherwise stated. Regime Model Sharpe Sortino Max Drawdown Turnover (ann.) COVID TFT+SVR 0.42 FinGPT 0.57 2022 TFT+SVR 0.36 FinGPT 0.44 2023 TFT+SVR 0.39 FinGPT 0.46 Low vol TFT+SVR 0.30 FinGPT 0.34 Mid vol TFT+SVR 0.34 FinGPT 0.41 High vol TFT+SVR 0.48 FinGPT 0.60 Open in a new tab Metrics are mean ± 95% CI via block bootstrap across folds and seeds (block length = 5 trading days). Ablations: drivers of improvement We quantify the contribution of ESG and ABSA feature blocks with fold-wise ablations. Removing either ESG or ABSA reduces return-forecasting accuracy and directional skill, indicating that both long-horizon quality (ESG) and short-horizon flow (ABSA) contribute complementary information. Table 11 reports the fold-wise ablation results, showing that removing either ESG or ABSA degrades forecasting accuracy and directional skill. Table 11. Ablations on return forecasting (aggregate across assets). Configuration MAE ( ) RMSE ( ) DA (%) Full (TFT+SVR, ESG+ABSA+Tech) 94.5 No ESG 90.2 No ABSA (no Sentiment) 90.8 Open in a new tab Values are mean ±95% CI across folds and seeds. ESG–sentiment interactions (quantified) Consistent with the gated fusion in Eqs. ( 1 )–( 3 ), we quantify cross–modal effects with SHAP interaction values and Friedman’s . For the TFT (a non–tree model), we apply KernelSHAP on the last-step inputs of each window: given , we explain the final step features . Per fold, we draw 500 background samples from the training fold (stratified by realized-volatility terciles) and compute the pairwise SHAP interaction matrix; the ESG–ABSA entry is summarized by its median absolute value across folds/assets. Friedman’s is computed on the same feature pair using the standard variance-decomposition estimator. Results show a statistically non-zero interaction across assets (median ; IQR ), with permutation tests (1,000 label permutations, BH-FDR 5%) rejecting in the majority of assets. To interpret the shape, we report 2D ALE on over a grid (centered effects; second-order ALE; light LOESS smoothing ). Qualitatively, negative ABSA combined with low ESG produces the largest downward adjustments, while high ESG partially mitigates negative sentiment. Regime-sliced SHAP confirms this: using 20-day realized-volatility terciles, ABSA importance rises in the top-tercile (high volatility) and ESG gains relative weight in the bottom-tercile (stable regimes). Robustness Stress windows. During COVID-19, the 2022 inflation shock, and the 2023 banking stress, TFT+SVR maintains strong risk-adjusted performance with stable error behavior. Figure 7 overlays predicted vs. actual paths with empirical bands computed from rolling 60-day residual standard deviations within each fold (per asset). These are not predictive quantiles from a probabilistic model; they visualize realized error dispersion around point forecasts. Figure 7. Open in a new tab Stress episode example with empirical residual bands. Price paths are reconstructed from predicted log returns via . SVR residual stabilization by regime is summarized in Table 12 , reporting variance reduction and skewness changes from TFT to TFT+SVR. Table 12. SVR residual stabilization by regime. Regime Var(TFT) Var(TFT+SVR) % Reduction Skew(TFT) Skew(TFT+SVR) COVID 2022 2023 Low vol Mid vol High vol Open in a new tab Residuals are daily return errors: . Var is residual variance; % Reduction is variance reduction from TFT to TFT+SVR; Skew is residual skewness. Variance-reduction estimates are computed fold-wise and summarized (block-bootstrap, 95% CI; omitted here for compactness). Note: Variance-reduction percentages are computed fold-wise and summarized using a block bootstrap (block length = 5 trading days). Confidence intervals are omitted for compactness Variance-reduction percentages were first computed fold-wise and then summarized across folds and seeds using a paired block bootstrap (block length 5 trading days); the resulting 95% confidence intervals (not shown for brevity) confirmed stable variance reduction, with the largest effects in high-volatility regimes. Interpretation. The residual corrector yields the strongest variance and tail-asymmetry reduction in high-volatility regimes, consistent with the role of SVR in correcting predictable residual structure under regime shifts. Simplified deployment. A one-layer/two-head TFT (no BiLSTM) reduces inference time to of the full model (i.e., reduction) (Table 13 ) while preserving the qualitative ranking and interaction patterns of the full model. Inference times are reported as relative units with the full model normalized to 100; measurements were taken on our test machine with a fixed batch size). Table 13. Latency-optimized variants (relative inference time; full 100). Variant Time Full (TFT+SVR) 100 TFT (2 layers, 4 heads) 63 TFT (1 layer, 2 heads) 55 Open in a new tab Discussion and limitations Interpretation. Across equities and crypto, the deployable default (TFT+SVR) delivers consistent, statistically significant gains over FinGPT on point and directional accuracy, and improves strategy-level Sharpe/Sortino. Improvements are modest per asset/metric but aggregate to economically meaningful deltas under conservative thresholds. Ablations indicate the gains arise from jointly leveraging short-horizon flow (ABSA) and slower-moving quality (ESG), fused through attention with an SVR residual corrector. Regime stability and mechanism. The regime-split analysis indicates that the observed improvements are not confined to a single market phase: TFT+SVR outperforms FinGPT during the COVID crash, the 2022 tightening cycle, and the 2023 banking stress window, and remains competitive in both low- and high-volatility terciles. Mechanistically, the regime-conditioned gate and SHAP interaction tests show a systematic shift toward sentiment features during turbulence and toward ESG in calmer regimes, aligning with financial intuition that short-horizon flow dominates during stress while resilience signals matter more in stable conditions. As expected, absolute forecast errors increase in high-volatility regimes for all models; however, TFT+SVR preserves a consistent relative advantage over the text-only baseline and shows lower drawdown and turnover under the same trading rule. Practicality vs. complexity. The recommended deployment omits BiLSTM: TFT+SVR is compact and latency-friendly while retaining of full-stack accuracy; the one-layer/two-head variant provides additional speed-up with limited accuracy loss. Threats to validity. Limitations include: (i) correlation vs. causation (we report predictive, not causal, interactions); (ii) training cost of rolling, multi-source features (mitigated by compact variants/compression); (iii) regime shifts not explicitly detected in real time; (iv) potential dilution from forward-filled low-frequency data (asynchronous fusion is future work); (v) event-window HAC-robust Diebold–Mariano tests in shorter stress episodes (e.g., 2023 banking stress) have lower statistical power, which is why we complement them with volatility-regime slicing and block-bootstrap confidence intervals; and (vi) backtest external validity (hence our emphasis on point and directional accuracy). Conclusion We presented a compact hybrid framework that couples a Temporal Fusion Transformer with a lightweight SVR residual corrector and an explicit, gated late-fusion of ESG and aspect-based sentiment signals. Evaluated under a finance-grade, leak-proof walk-forward protocol across equities and crypto (2020–2024), the approach delivers statistically significant gains in point and directional accuracy and translates these into higher risk-adjusted performance relative to strong baselines. The deployable variant—kept intentionally small—preserves most accuracy benefits at substantially lower latency, supporting real-time or near-real-time use. Regime-split evaluation confirms that these gains persist across major market cycles (COVID shock, 2022 tightening, 2023 banking stress) and across volatility-defined regimes. Mechanism diagnostics (gate dynamics, regime-sliced SHAP interaction shifts, and residual stabilization) provide evidence that the model adapts its reliance on ESG versus sentiment in a regime-consistent manner, improving robustness without sacrificing interpretability. Beyond headline metrics, the framework remains interpretable: SHAP interactions and ALE analyses show that sentiment predominates in high-volatility regimes while ESG contributes more in stable conditions, with their interaction quantitatively non-zero. These insights are predictive rather than causal but provide actionable guidance for regime-aware weighting. Limitations. First, mixed-frequency alignment (e.g., quarterly ESG vs. daily prices) required conservative forward-filling, which may dilute signals. Second, regime shifts are detected implicitly via the model rather than by explicit change-point logic. Third, economic results come from simplified, uniform-friction backtests; alternative cost models and execution constraints may alter absolute levels. Future work. We will (i) incorporate lightweight regime detectors to adapt horizons and hyperparameters on the fly, (ii) replace forward-filling with asynchronous/state-space fusion for mixed-frequency data, (iii) distill the sentiment pipeline and explore model compression to further reduce latency, and (iv) expand robustness tests to heterogeneous transaction-cost and slippage models, as well as prospective live evaluation. Taken together, these steps aim to enhance deployability without sacrificing interpretability or rigor. Author contributions S.M. contributed to financial interpretation, benchmarking design, and critical revision of the manuscript for quantitative-finance relevance. Z.L.M. conceived the study, developed the methodology and software, conducted the experiments, and performed the statistical analysis. C.Y.L. curated and harmonized the datasets (market, ESG, and sentiment), implemented the leak-proof evaluation protocol, and validated the experimental results. P.K.S. contributed to methodological review, robustness analysis, and manuscript editing, with emphasis on experimental rigor and reproducibility. A.K.S. contributed to validation, presentation of results, and manuscript refinement, including clarity of the deployment and latency analysis. All authors reviewed and approved the final manuscript. Funding Open access funding provided by Manipal Academy of Higher Education, Manipal. No funds, grants, or other financial support were received for this study. Data availability Daily prices for equities/indices (Yahoo Finance) and crypto (Binance API) are publicly available. Macroeconomic series are obtained from World Bank/IMF portals. The FiQA corpus is publicly available for research use. ESG metrics were accessed via Bloomberg ESG under a license and cannot be redistributed. To support reproducibility without releasing proprietary fields, we provide fold-wise aligned and standardized feature matrices (with proprietary ESG columns redacted), walk-forward split files, and experiment manifests sufficient to reproduce all headline tables and figures. Code availability Code to reproduce the experiments is provided as a containerized artifact including a pinned environment.yml, Dockerfile, fold-wise split CSVs, fixed seed configuration, and scripts that regenerate all tables and figures. Proprietary ESG fields are excluded; instructions and public-data stubs are included to enable replication with open sources. Declarations Competing interests The authors declare no competing interests. Ethics approval and consent to participate Not applicable. The work uses secondary market, macroeconomic, and vendor datasets; no human or animal subjects were involved. Use of AI tools During the preparation of this manuscript, the authors used generative AI tools (including OpenAI ChatGPT) solely for language editing (grammar, spelling, and readability). The authors reviewed and edited all outputs and take full responsibility for the content of the manuscript. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Contributor Information Zefree Lazarus Mayaluri, Email: [email protected]. Prabodh Kumar Sahoo, Email: [email protected]. Aswini Kumar Samantaray, Email: [email protected]. References 1. Godfrey, M. D., Granger, C. W. J. & Morgenstern, O. The random-walk hypothesis of stock market behavior. Kyklos 17 , 1–30. 10.1111/j.1467-6435.1964.tb02458.x (1964). [ Google Scholar ] 2. Jensen, M. C. Some anomalous evidence regarding market efficiency. J. Financ. Econ. 6 , 95–101 (1978). [ Google Scholar ] 3. Fama, E. F. Random walks in stock market prices. Financ. Anal. J. 51 , 75–80. 10.2469/faj.v51.n1.1861 (1995). [ Google Scholar ] 4. Ghallabi, F., Souissi, B., Du, A. M. & Ali, S. Esg stock markets and clean energy prices prediction: Insights from advanced machine learning. Int. Rev. Financ. Anal. 97 , 103889 (2025). [ Google Scholar ] 5. Su, L. et al. A systematic review for transformer-based long-term series forecasting. Artif. Intell. Rev. 58 , 80. 10.1007/s10462-024-11044-2 (2025). [ Google Scholar ] 6. Khalil, M. A., Khalil, S. & Sinliamthong, P. From ratings to resilience: The role and implications of environmental, social, and governance (esg) performance in corporate solvency. Sustainable Futures 8 , 100304 (2024). [ Google Scholar ] 7. Li, X. & Yuan, J. Deeptvar: Deep learning for a time-varying var model with extension to integrated var. Int. J. Forecast. 40 , 1123–1133 (2024). [ Google Scholar ] 8. Graves, A. Long Short-Term Memory, 37–45 (Springer, Berlin Heidelberg, Berlin. Heidelberg 10.1007/978-3-642-24797-2_4 (2012). [ Google Scholar ] 9. Das, S. S., Mishra, S., Mayaluri, Z. L. & Panda, G. Dependable and secure ai-driven fintech adoption for rural tourism & entrepreneurship in odisha: A cyber-physical systems perspective. SN Comput. Sci. 6 , 439. 10.1007/s42979-025-03995-2 (2025). [ Google Scholar ] 10. Han, Y., Dutta, B., García-Zamora, D. & Martínez, L. Robust comprehensive minimum cost consensus model for multi-criteria group decision making: Application in iot platform selection. Adv. Complex Decis. Mak. 88–108 (2023). 11. de Jesus, D. P., de Medeiros, E. H. O., Godeiro, L. L. & Proque, A. L. Forecasting Brazilian stock market using sentiment indices from textual data, chat-gpt-based and technical indicators. Comput. Econ. 10.1007/s10614-024-10835-7 (2025) ( In press ). [ Google Scholar ] 12. Smola, A. J. & Schölkopf, B. A tutorial on support vector regression. Stat. Comput. 14 , 199–222. 10.1023/B:STCO.0000035301.49549.88 (2004). [ Google Scholar ] 13. Mattera, R., Athanasopoulos, G. & Hyndman, R. Improving out-of-sample forecasts of stock price indexes with forecast reconciliation and clustering. Quant. Financ. 24 , 1641–1667. 10.1080/14697688.2024.2412687 (2024). [ Google Scholar ] 14. Krishnamurthy, S., Adewuyi, O. B., Luwaca, E., Ratshitanga, M. & Moodley, P. Artificial intelligence-based forecasting models for integrated energy system management planning: An exploration of the prospects for south africa. Energy Convers. Manag. X 24 , 100772 (2024) https://www.sciencedirect.com/science/article/pii/S2590174524002502. . [ Google Scholar ] 15. Dudek, G. Std: A seasonal-trend-dispersion decomposition of time series. IEEE Trans. Knowl. Data Eng. 35 , 10339–10350 (2023). [ Google Scholar ] 16. Chiang, W.-C., Urban, T. & Baldridge, G. A neural network approach to mutual fund net asset value forecasting. Omega 24 , 205–215 (1996) https://www.sciencedirect.com/science/article/pii/0305048395000593. . [ Google Scholar ] 17. Cortes, C. & Vapnik, V. Support-vector networks. Mach. Learn. 20 , 273–297. 10.1007/BF00994018 (1995). [ Google Scholar ] 18. Trafalis, T. B. & Ince, H. Support vector machine for regression and applications to financial forecasting. In Proceedings of the IEEE-INNS-ENNS International Joint Conference on Neural Networks (IJCNN 2000) 6, 348–353 (2000). 19. Gupta, S., Kumar, V., Ahalawat, K. & Sagar, V. Artificial intelligence role in stock market: A comprehensive analysis. Challenges in Information, Communication and Computing Technology 6–15 (2024). 10.1201/9781003559085-108. 20. Jae Kim, K. Financial time series forecasting using support vector machines. Neurocomputing 55 , 307–319 (2003). 21. Safari, A. & Ghaemi, S. Neurofuzzyman: A hybrid neuro-fuzzy bilstm stacked ensemble model for financial forecasting and analysis: Dataset case studies on jpmorgan, amzn and tsla. Expert Syst. Appl. 266 , 126037 (2025) https://www.sciencedirect.com/science/article/pii/S095741742402904X. . [ Google Scholar ] 22. Ge, Q. Enhancing stock market forecasting: A hybrid model for accurate prediction of s&p 500 and csi 300 future prices. Expert Syst. Appl. 260 , 125380 (2025) https://www.sciencedirect.com/science/article/pii/S0957417424022474. . [ Google Scholar ] 23. Patel, J., Shah, S., Thakkar, P. & Kotecha, K. Predicting stock market index using fusion of machine learning techniques. Expert Syst. Appl. 42 , 2162–2172 (2015) https://www.sciencedirect.com/science/article/pii/S0957417414006551. . [ Google Scholar ] 24. Li, J., Chen, W., Zhou, Z., Yang, J. & Zeng, D. Deepar-attention probabilistic prediction for stock price series. Neural Comput. Appl. 36 , 15389–15406. 10.1007/s00521-024-09916-3 (2024). [ Google Scholar ] 25. Alves Portela Santos, A., Carneiro Affonso da Costa, N. & dos Santos Coelho, L. Computational intelligence approaches and linear models in case studies of forecasting exchange rates. Expert Syst. Appl. 33, 816–823 (2007). 26. Araci, D. Finbert: Financial sentiment analysis with pre-trained language models (2019). arXiv:1908.10063 . 27. Liu, Z., Huang, D., Huang, K., Li, Z. & Zhao, J. Finbert: A pre-trained financial language representation model for financial text mining. Proceedings of the Twenty-Ninth International Joint Conference on Artificial Intelligence (IJCAI-21) 4513–4519 (2021). https://www.ijcai.org/proceedings/2020/0622.pdf . 28. Ağır, T. T. Daily global solar radiation prediction with hybrid lstm-svm: The case of nusaybin. Arab. J. Sci. Eng. 10.1007/s13369-025-10322-7 (2025). [ Google Scholar ] 29. Wang, X. et al. Ecoforecast: An interpretable data-driven approach for short-term macroeconomic forecasting using n-beats neural network. Eng. Appl. Artif. Intell. 114 , 105072 (2022) https://www.sciencedirect.com/science/article/pii/S0952197622002299. . [ Google Scholar ] 30. Ke, Z. et al. A stable technical feature with gru-cnn-ga fusion. Appl. Soft Comput. 187 , 114302 (2026) https://www.sciencedirect.com/science/article/pii/S1568494625016151. . [ Google Scholar ] 31. Kehinde, T. O. et al. Helformer: an attention-based deep learning model for cryptocurrency price forecasting. J. Big Data 12 , 81. 10.1186/s40537-025-01135-4 (2025). [ Google Scholar ] 32. Bergmeir, C. Common pitfalls and better practices in forecast evaluation for data scientists. Foresight Int. J. Appl. Forecast. 70 , 5–12 (2023). [ DOI ] [ PMC free article ] [ PubMed ] 33. Hewamalage, H., Ackermann, K. & Bergmeir, C. Forecast evaluation for data scientists: common pitfalls and best practices. Data Min. Knowl. Disc. 37 , 788–832. 10.1007/s10618-022-00894-5 (2023). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 34. Anatolyev, S. & Gerko, A. A trading approach to testing for predictability. J. Bus. Econ. Stat. 23 , 455–461. 10.1198/073500104000000640 (2005). [ Google Scholar ] 35. Diebold, F. X. & Mariano, R. S. Comparing predictive accuracy. J. Bus. Econ. Stat. 20 , 134–144. 10.1198/073500102753410444 (2002). [ Google Scholar ] 36. Vuletić, M., Prenzel, F. & Cucuringu, M. Fin-gan: forecasting and classifying financial time series via generative adversarial networks. Quant. Financ. 24 , 175–199. 10.1080/14697688.2023.2299466 (2024). [ Google Scholar ] 37. Xing, F. Z., Cambria, E. & Welsch, R. E. Natural language based financial forecasting: a survey. Artif. Intell. Rev. 50 , 49–73. 10.1007/s10462-017-9588-9 (2018). [ Google Scholar ] 38. Banerjee, S., Aggarwal, D. & Sengupta, P. Do stock markets care about esg and sentiments? impact of esg and investors’ sentiment on share price prediction using machine learning. Ann. Oper. Res. 10.1007/s10479-025-06480-4 (2025). [ Google Scholar ] 39. Salvagnin, C., Glielmo, A., Giuli, M. E. D. & Mira, A. Investigating the price determinants of the european emission trading system: a non-parametric approach. Quant. Financ. 24 , 1529–1544. 10.1080/14697688.2024.2407895 (2024). [ Google Scholar ] 40. Neely, C. J., Rapach, D. E., Tu, J. & Zhou, G. Forecasting the equity risk premium: The role of technical indicators. Manage. Sci. 60 , 1772–1791 (2014) http://www.jstor.org/stable/42919633. . [ Google Scholar ] 41. Hyndman, R. & Athanasopoulos, G. Forecasting: Principles and Practice 3rd edn. (OTexts, Australia, 2021). [ Google Scholar ] 42. Bai, J. & Ng, S. Determining the number of primitive shocks in factor models. J. Bus. Econ. Stat. 25 , 52–60. 10.1198/073500106000000413 (2007). [ Google Scholar ] 43. Stock, J. H. & Watson, M. W. Forecasting using principal components from a large number of predictors. J. Am. Stat. Assoc. 97 , 1167–1179 (2002) http://www.jstor.org/stable/3085839. . [ Google Scholar ] 44. Phuoc, T., Anh, P. T. K., Tam, P. H. & Nguyen, C. V. Applying machine learning algorithms to predict the stock price trend in the stock market – the case of vietnam. Hum. Soc. Sci. Commun. 11 , 393. 10.1057/s41599-024-02807-x (2024). [ Google Scholar ] 45. Lu, X., Poon, J. & Khushi, M. Leveraging bilstm-gat for enhanced stock market prediction: a dual-graph approach to portfolio optimization. Appl. Intell. 55 , 601. 10.1007/s10489-025-06462-w (2025). [ Google Scholar ] 46. Yang, H., Liu, X.-Y. & Wang, C. D. Fingpt: Open-source financial large language models (2023). arXiv:2306.06031 . Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Data Availability Statement Daily prices for equities/indices (Yahoo Finance) and crypto (Binance API) are publicly available. Macroeconomic series are obtained from World Bank/IMF portals. The FiQA corpus is publicly available for research use. ESG metrics were accessed via Bloomberg ESG under a license and cannot be redistributed. To support reproducibility without releasing proprietary fields, we provide fold-wise aligned and standardized feature matrices (with proprietary ESG columns redacted), walk-forward split files, and experiment manifests sufficient to reproduce all headline tables and figures. Code to reproduce the experiments is provided as a containerized artifact including a pinned environment.yml, Dockerfile, fold-wise split CSVs, fixed seed configuration, and scripts that regenerate all tables and figures. Proprietary ESG fields are excluded; instructions and public-data stubs are included to enable replication with open sources. Articles from Scientific Reports are provided here courtesy of Nature Publishing Group ACTIONS View on publisher site PDF (4.8 MB) Cite Collections Permalink PERMALINK Copy RESOURCES Similar articles Cited by other articles Links to NCBI Databases Cite Copy Download .nbib .nbib Format: AMA APA MLA NLM Add to Collections Create a new collection Add to an existing collection Name your collection * Choose a collection Unable to load your collection due to an error Please try again Add Cancel Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top