Contemporary information technology applications are overwhelmed by big data and require advanced data science analytics for careful investigation, interpretation, and predictions. Data sets in various applications exhibit high frequency with non-linear dynamic variability, and hence, leveraging the strengths of data-driven feature selection and sophisticated machine learning architectures becomes essential.This study proposes two novel architectures- Data-Driven Long Short-Term Memory (DD-LSTM) and Data-Driven Gated Recurrent Unit (DD-GRU), to improve predictive accuracy for highly fluctuating time-series data. Input data are log-transformed and used to derive data-driven risk forecasts and non-linear residuals based on underlying statistical features, which are then integrated with normalized original data into hyperparameter-optimized LSTM and GRU models. Experimental results with a financial dataset show that the proposed frameworks significantly outperform conventional LSTM and GRU by capturing intricate temporal patterns and risk dynamics. DD-GRU, in particular, exhibits greater computational efficiency, making it a robust solution for modeling nonlinear and irregular time-series data. This research not only addresses the critical challenge of optimizing temporal feature selection in high-frequency datasets but also offers a robust framework for analyzing complex temporal patterns across diverse high-frequency data sources.

Forecasting stock market indices is a challenging task due to the inherent complexity, non-linearity, and stochastic nature of financial time series. Although deep learning models, such as Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) architectures, outperform traditional econometric methods in capturing temporal dependencies, their standalone implementations often lack robustness and generalizability. This study presents a novel hybrid framework combining recurrent neural networks with Support Vector Regression (SVR) to address these limitations. The proposed approach integrates the sequential learning capabilities of LSTM and GRU with the nonlinear regression strengths of SVR, achieving superior predictive performance across diverse global indices. Empirical results highlight significant improvements over baseline models in both accuracy and adaptability to varying market conditions. The hybrid framework demonstrates its effectiveness in merging the advantages of its components, providing robust, generalizable predictions with reduced susceptibility to overfitting. These findings pave the way for future research in hybrid financial forecasting methods and their practical applications.

This study investigates data-driven financial correlation networks with a particular focus on modeling log returns using a data-driven t-distribution. A key contribution is showing that the covariance matrix of absolute log returns, used in volatility networks can be expressed as a function of the sign correlation and the covariance matrix of log returns. Fuzzy adjacency matrices are introduced to account for the uncertainty inherent in correlation estimates that depend on the degrees of freedom. Using data from 55 stocks across 13 sectors and S&P500 (big data), we compare four different networks based on correlations of log returns, volatility of log returns, data-driven volatility of log returns and sign(±) of returns. Three dynamic networks based on price correlations, ARIMA innovation correlations of the price and neuro innovation correlations of the price are also considered. All seven networks are evaluated using metrics such as average node degree, sparsity, and maximum eigenvalue. Unlike the existing work, the novelty of this study is demonstrating a relationship between data driven volatility networks and recently proposed volatility correlation networks without using distributional properties. Three different community detection algorithms demonstrate that the proposed data-driven volatility network proposed in this paper has a higher modularity score.

The eigenvalues of data-driven exponentially weighted moving average (DDEWMA) covariance matrix of stock returns plays an important role in portfolio optimization. Moreover, the eigenvalues of the Laplacian matrix play a fundamental role in financial network analysis. The novelty of this paper is to apply Random Matrix Theory (RMT) to denoise three covariance matrices ( empirical covariance, neuro covariance and DDEWMA covariance). Simulation study shows that portfolio weights generated using RMT filtered empirical covariance matrices based on both normal and t-distributed returns are more resilient. The networks built from t-distributed returns are more resilient (fewer but stronger connections with smaller average degree of nodes) than the networks built from normally distributed returns. Unlike the existing work, the novelty of this paper is to first show that the superiority of the RMT-filtered DDEWMA covariance model in portfolio optimization and, then show that the networks built from t-distributed returns are more resilient.

Forecasting stock prices and volatility plays a crucial role in making better investment decisions. Recently, there has been a growing interest in studying the Forecast Linear Augmented Projection (FLAP) method for multivariate normally distributed time series to reduce the forecast error variance and mean squared error (MSE) of long-term forecasts. FLAP method consists of three steps: formation of principal components, forecasting original and component series (base forecasts), and then projecting those base forecasts. First, this paper shows that the non-linear prediction (conditional mean) has a smaller MSE than the linear prediction. Then, it demonstrates that for multivariate time series, forecasts using the neural network autoregressive (NNAR) model with FLAP outperform those obtained by the commonly used autoregressive integrated moving average (ARIMA) model forecasts with FLAP (resulting in lower out-of-sample root mean squared error (RMSE) in general). For simulated multivariate data with t-distributed errors, the combined forecast model with NNAR and FLAP outperforms the NNAR forecast model. In the empirical study, we forecast stock price and volatility for 55 most traded stocks and S&P500 index using the ARIMA model with FLAP and the NNAR model with FLAP. The results show that FLAP is more effective in reducing forecast error variance and RMSE for non-stationary multivariate price and volatility forecasts.

Recently, there has been a growing interest in using neuro volatility models in fuzzy forecasting and fuzzy option pricing. Neuro volatility models are used to model and predict financial market volatility by extending the neural network autoregressive (NNAR) model for nonlinear nonstationary times series data. In financial risk forecasting, various risk forecasting models for volatility are used to obtain the volatility forecasts, and the model risk ratio based on all the models is calculated to assess the stability of the financial system. However, the recently proposed neuro volatility models (based on neural networks such as LSTM, NNAR, etc.) are not used in evaluating the model risk. In this paper, novel 'neuro model risk forecasts' are obtained by including recently proposed neuro volatility models, and the resiliency of the financial system is studied. Unlike the existing model risk ratio forecasting based on linear volatility models, the driving idea is to use more appropriate nonlinear nonstationary neuro volatility forecasting models to obtain the model risk forecasts. The proposed model risk forecasts in this paper have been evaluated through extensive experiments, and it is shown that the model risk ratio can effectively serve as a metric for assessing the resilience and stability of the targeted financial system.

Recently, there has been a growing interest in utilizing empirical correlations of log returns to examine financial network models for stock prices by representing stocks as nodes and their relationships as edges. In this paper, empirical cor-relation, data-driven correlation, and cosine similarity matrices are used to obtain adjacency matrices for connectedness. For a given number of nodes, the simplest Erdos and Rényi (ER) model can be obtained by fixing the number of stocks and estimating the probability for any two vertices to be connected by an edge. Unlike existing work, the driving idea in this paper is to estimate the probability with the median cross-correlation to construct the ER network for observed volatility. Moreover, to address the uncertainty associated with these estimates of connectedness, this study introduces data-driven non-linear adaptive symmetric fuzzy adjacency matrices. In the literature, a certain thresh-old is identified for a network considering its connectedness. In this study, threshold and fuzzy parameters for the corre-sponding minimally connected fuzzy networks are determined, revealing unique network structures and the most significant stocks/cryptocurrencies (nodes with the highest number of links). Furthermore, as an application of the suggested fuzzy networks, three clustering approaches, fuzzy network clustering, k-means clustering, and Sharpe ratio clustering, are applied followed by the PageRank algorithm to construct optimal portfolios. A comparison of the constructed portfolios suggests that the fuzzy networks and clustering techniques are capable of yielding high cumulative returns during the study period.