ROBUST QUANTILE REGRESSION-BASED MACHINE LEARNING FRAMEWORK FOR OUTLIER-RESILIENT TIME SERIES ANALYSIS

Journal of Theoretical and Applied Information Technology 31st May 2025. Vol.103. No.10 © Little Lion Scientific ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195 4414 ROBUST QUANTILE REGRESSION-BASED MACHINE LEARNING FRAMEWORK FOR OUTLIER-RESILIENT TIME SERIES ANALYSIS Mr. TUSHAR MEHTA1 , DR.DHARMENDRA PATEL2 1,2 Smt. Chandaben Mohanbhai Patel Institute of Computer Applications, Charotar University of Science and Technology Changa, Gujarat, India. E-mail: [email protected], [email protected] ABSTRACT Time series analysis is a powerful tool in countless regions, from finance to healthcare, but is often challenged by the presence of outliers that can distort predictions and model output. This article presents a robust quantile regression-based framework for machine learning, which increases the resilience of time series analysis compared to outliers. By using quantile regression, the proposed framework captures the conditional distribution of time series data and provides a more comprehensive understanding of its underlying structure. Machine learning integration improves the ability of models to adapt to complex nonlinear patterns while simultaneously maintaining robustness to anomaly data points. Through extensive research into synthesis and practical datasets, we show that our framework outweighs traditional predictability and trigger resilience methods. The results highlight the possibility of combining quantile regression with machine learning for robust time series analysis, providing promising directions for future research and applications in the environment. Time series forecasts play a key role, especially in financial markets where accurate forecasts are useful for investors and stakeholders. However, traditional models have difficulty recording nonlinear dependencies and are not able to effectively handle outliers. This article presents a robust quantile regression-based machine learning framework for diffusion-preserving time series analysis. It integrates long-term time memory (LSTM), LightGBM, and stacked ensemble models. The proposed ensemble approach uses quantile regression for robust outsourcing processing while combining deep learning strengths to increase base techniques to improve predictive performance. Experimental evaluation of Goldman Sachs BDC, Inc.(GSBD) shared course data demonstrates the advantages of the ensemble model compared to the individual model. The results show that the stacked ensemble model reaches the lowest flipper loss (0.0656), MAE (0.1313), RMSE (0.2185), and the highest R² (0.9778), exceeding LSTM and LightGBM. The results highlight the effectiveness of hybrid ensemble learning in financial series forecasting, providing a more accurate and robust approach to dealing with outlier sacrificial data. Keywords: Time Series Analysis, Quantile Regression, Outlier Resilience, Machine Learning Framework, Robust Prediction. 1. INTRODUCTION In areas such as banking, health, energy, and climate research, time series analysis is the pillar of decision-making. Forecasting and strategic planning rely heavily on the ability to spot trends, patterns, and anomalies in continuous data. Nevertheless, traditional time series models are seriously questioned by the presence of outliers and the presence of extreme or unusual data values. Especially in situations where data integrity is very important, outlier predictions can be skewed, reducing the output of the model and leading to false conclusions. Therefore, there is a high demand for powerful technologies, both in research and practical terms, that can be managed efficiently while maintaining high prediction accuracy at the same time. Machine learning methods (ML) have been a more effective tool for time series analysis in recent years, as it provides the adaptability of complex, nonlinear interactions. It is known that their longterm memory (LSTM) networks for their ability to recognize temporal dependencies of continuous data are known. LGBM models (light gradient boosting machines) use structured data to demonstrate outstanding skills and produce Journal of Theoretical and Applied Information Technology 31st May 2025. Vol.103. No.10 © Little Lion Scientific ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195 4415 interpretable results. These models are powerful, but of course, these models are not intended to solve the problems presented by outliers. This limitation highlights the need for hybrid systems that combine the benefits of modern machinelearning approaches with statistical methods. Quantile regression, a powerful statistical method, provides a possible answer to this question. Quantile regression models the conditional quantum of response variables, in contrast to traditional regression techniques, highlighting the mean of the dependent variable. This feature is particularly suitable for research without research, as it provides a more comprehensive knowledge of data distributions that include tails. The use of the properties of both paradigms, the robustness of quantile regression, and the predictability of machine learning through the combination of quantile regression and machine learning models, including LSTM and LGBM is practical. This work beats a unique and robust quantile regression-based framework for machine learning before series analysis using outliers. LSTM, LGBM, and ensemble hybrid methods are combined in context to improve the accuracy and robustness of predictions against outliers. The LGBM model pays for structured functionality management and interpretable knowledge specifications, while the LSTM components record temporary patterns and long-term dependencies of the data. The combined method guarantees resilience to outliers by combining these models with quantile regression to improve general performance. Through thorough experiments on synthesis and actual data records, we demonstrate the advantages of the proposed framework compared to traditional approaches on predictive accuracy and outliers. In Section 2, relatives work in time series analysis. The entire publication discusses quantile regression and machine learning. Section 3 describes the proposed architecture, including ensemble methods, LSTM, and LGBM integration. Section 4 provides experimental designs, data records, and results. Section 5 concludes the work with a discussion of results and potential research directions. This work aims to open paths for more reliable and accurate time series analysis in a context that tends to bridge the gap between statistical approaches and advanced technologies for machine learning. 2. LITERATURE REVIEW Jiang, Z et al. [1] develops Centra-Net to optimize visual locations with a variety of data records, simultaneously maintaining storage efficiency. There is a basic feature extraction unit (BFEU) with parallel branching for local property extraction and attention-based calibration. The Filter Company Sharing Mechanism (FSM) adaptively manages parameter releases, while the Complexity Promotional Algorithm (CPGA) compensates for learning based on task complexity. Large-scale experiments show that Centra-net surpasses cuttingedge models with fewer parameters. Gong, J et al. [2] proposes a hybrid ensemble model combining TCN, LSTM, and LightGBM to combine the model costs using several linear regressions. The ensemble model is evaluated on Australian and Chinese data records and reduces MAE by 4.88% and 28.95% compared to TCN-LSTM or LightGBM, beyond the individual models. The proposed multiple linear regression-ensemble method continues to improve by 3.64% and RMSE. This demonstrates its effectiveness in improving predictive performance. Zhao, T et al. [3] suggest improving prediction accuracy by using the ability of LightGBM to efficiently deal with nonlinear relationships and prevent over-management while also using the ability of TCN to capture seasonal trends. It suggests the LIGHTGB model. Experimental results show that TCN-LightGBM surpassed the conventional model, achieving 91.3°C narration with an MSE of 0.021 and MAE of 0.115. Additionally, parameter sensitivity analysis confirms the robustness of the model and provides a reliable means of predicting agricultural trade and other forecast tasks in time series. Maharina, M et al. [4] address the challenges of flood prediction in Jakarta, Indonesia using machine learning models (ML) to improve the accuracy of forecasts using weather data. As traditional models combat Jakarta's highly diverse climate, this study evaluates logistics regression (LR), LightGBM and XGBoost based on five performance metrics. The proposed approach includes data preprocessing, functional engineering, and overloading with cross-validation. This greatly improves performance. Results show that LightGBM reached 96.81°Curacy, Xgboost 96.67%, and LR 93.82%, indicating the effectiveness of the model in flood risk management and damage limits. Zhang, C et al.[5] proposes a form-supported EE-LightBM model that integrates sub-techniques and hybrid ensemble strategies to improve learning from limited error samples. The SHAP method enhances Journal of Theoretical and Applied Information Technology 31st May 2025. Vol.103. No.10 © Little Lion Scientific ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195 4416 explainability by identifying key features contributing to faults. Evaluation of real-world network data shows high detection accuracy (0.9968), F1-score (0.9711), and low false negative (0.0033) and false positive (0.0032) rates, demonstrating the model’s effectiveness in fault diagnosis. Airlangga, G. [6] explores machine learning techniques for tsunami prediction using seismic data from 1995 to 2023. Various models, including ensemble methods (Random Forest, Gradient Boosting, XGBoost, LightGBM, CatBoost) and Neural Networks, were evaluated using 10-fold cross-validation. Gradient Boosting achieved the highest ROC-AUC (0.8606) and a balanced F1-score (0.6544), while Random Forest excelled in precision (0.6920). Neural Networks underperformed, highlighting the need for further optimization. The study confirms that ensemble methods, particularly Gradient Boosting and Random Forest, are effective for tsunami prediction, aiding early warning systems. Future improvements will focus on enhancing recall and exploring hybrid models. Kumar, S et al. [7] present a hybrid prediction model that integrates LightGBM, XGBoost, Lasso regression, and random forests. This model surpasses basic line regressors over some valuation metrics, showing excellent accuracy and robustness. The research uncovers the possibilities of hybrid models in predictive analytics and proposes future inspections of additional algorithms and optimization techniques for further improvement. Kumar, M. S et al. [8] introduce an advanced ensemble stacking approach that summarises LightGBM, Plsregrechente, and ElaticNet to improve predictive modeling. LightGBM guarantees efficiency, PLSREGECHEN helps in feature selection, and ElasticNet guarantees regularization to prevent overhanging. Empirical results show superior performance compared to individual models, indicating the potential of this ensemble method to improve predictability in complex, real-world applications. Hola, A et al. [9] examine predictions (PDM) using IoT sensor data and machine learning to predict equipment failures, optimize maintenance plans, and reduce downtime. Compare regression models including LSTM, XGBoost, and LightGBM based on RMSE and CPU usage by analyzing CMAPSS data records from NASA. This study concludes that LSTM represents optimal performance for predicting useful life (RULL) and demonstrates its effectiveness in predictive expectations of aircraft engines. Sui, M et al. [10] study enhances retail investors' stock price predictions by integrating deep learning and time series models. Using techniques such as DNN, LightGBM, LSTM, GRU, and linear regression, the proposed ensemble model improves accuracy by preprocessing data, functional engineering, and adjusting hyperparameters. The results show that the ensemble approach goes beyond the individual models and provides a robust forecasting solution for retail investors. Examines approaches to machine learning to predict harmful algae blooms (HABs). Ahn, J. M et al. [11] propose a significant risk to ecosystems and human health. Compare gradient boosting models (XGBoost, LightGBM, CatBoost) with note-based CNN-LSTM models. It uses Bayesian optimization, and back and stacking techniques to improve accuracy. The results show that ensemble techniques improve predictive power by easing the limitations of individual models and highlighting the possibilities of AI-controlled prediction in HABS management. Gupta, V., & Kumar, E. [12] suggest an optimized LightGBM model.This is improved with the Harris Hawk Optimization (H3O) H3O algorithm for prediction. By using exclusive bundles of features (EFB) and gradient-based single-sided samples (GOSD), the model reduces over-adaptation and improves prediction accuracy. Furthermore, the Markov chain model improves residual errors, leading to improved stock trading and economic prognosis decisions. Experimental results demonstrate the effectiveness of this hybrid approach in predicting market trends and economic growth with high accuracy. Hartanto, A. D. et al. [13] propose a prediction model using LightGBM models (Light Gradient Boosting Machine), known for their accuracy and efficiency. By optimizing hyperparameters with Grid Search Cross Validation (GSCV), LightGBM studies are compared to other boost models such as Xgboost, Adaboost, and CatBoost. The experimental results show that LightGBM exceeds its counterparts, highlighting the potential for stock market forecasts. Future work will show that data preprocessing will be improved, external variables will be improved, and prediction accuracy will be improved.Kumar, S[14] proposes a framework for accurate prediction of the final price of Indian investment funds using an automated Arimax model. This compares the performance of LightGBM with Auto-Arimax and Facebook Prophet on the same data set. This indicates that the proposed model exceeds the alternative. Evaluations based on MSE, RMSE, and MAPE confirm excellent accuracy. Future work will show that additional parameters can be included to improve time series forecasting. Kumar, S et al. [15] present a new hybrid ensemble learning Journal of Theoretical and Applied Information Technology 31st May 2025. Vol.103. No.10 © Little Lion Scientific ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195 4417 approach that uses a stack of regressors to improve price forecasts for investment funds in Indian financial markets. Traditional forecasting methods struggle to capture complex nonlinear relationships in financial data, which the proposed ensemble model addresses by leveraging multiple base learners. Ridge regression is employed as the metaregressor, achieving superior predictive accuracy with metrics such as an R² score of 0.99998 and a low RMSE of 0.0041. The study demonstrates that the stacking-based ensemble model outperforms individual models, making it valuable for portfolio optimization and risk reduction. 2.1 Objectives 1.Develop a robust quantile regression-based machine learning framework capable of effectively modelling time series data in the presence of outliers. 2.Enhance the accuracy and reliability of forecasts by integrating multiple models and leveraging ensemble techniques. 3.Rigorously evaluate the proposed framework against traditional methods using real-world datasets and comprehensive error metrics. 4.Provide a scalable and interpretable tool to support decision-making in critical applications such as financial forecasting and resource management. 3. Research Methodologies Figure 1. Database Diagram Figure 2. The Architecture of the Proposed Model Journal of Theoretical and Applied Information Technology 31st May 2025. Vol.103. No.10 © Little Lion Scientific ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195 4418 3.1 Data sample GSBG Stock Data S. N Date Close High Low Open Volume 0 2018-01-02 10.6592931 74743652 10.678150368169 85 10.4754303684189 7 10.546147316576 42 181100 1 2018-01-03 10.5367136 00158691 10.701718240720 147 10.5037128518869 91 10.631001323905 988 111500 2 2018-01-04 10.5367136 00158691 10.619686653183 84 10.5136130763685 10.531999079090 884 76900 3 2018-01-05 10.3528528 21350098 10.541430102697 431 10.2868513124004 03 10.541430102697 431 161300 4 2018-01-08 10.2727069 85473633 10.414139024422 864 10.1737038388467 93 10.414139024422 864 204700 177 6 2025-01-27 12.8100004 196167 12.960000038146 973 12.7399997711181 64 12.800000190734 863 1215600 177 7 2025-01-28 12.8599996 56677246 12.970000267028 809 12.7799997329711 91 12.779999732971 191 961200 177 8 2025-01-29 12.6099996 56677246 12.899000167846 68 12.5600004196167 12.859999656677 246 657900 177 9 2025-01-30 12.7200002 67028809 12.770000457763 672 12.6049995422363 28 12.710000038146 973 734800 178 0 2025-01-31 12.8400001 5258789 12.850000381469 727 12.6949996948242 19 12.729999542236 328 847500 Journal of Theoretical and Applied Information Technology 31st May 2025. Vol.103. No.10 © Little Lion Scientific ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195 4419 LSTM Below is an overview of the mathematical foundation underlying Long Short-Term Memory (LSTM) networks. LSTM is a type of recurrent neural network (RNN) specifically designed to address the vanishing (and exploding) gradient problems encountered when learning long-term dependencies. The key innovation in LSTM is the introduction of a memory cell and gating mechanisms that regulate the flow of information. Consider an LSTM unit at the time step with input vector previous hidden state , and previous cell state . The following equations. describe the operations within an LSTM cell: Forget Gate The forget gate determines which information to discard from the previous cell state. (1) where is the logistic sigmoid function, is the weight matrix for the forget gate, and is the bias vector. Input Gate and Candidate Cell State: The input gate controls which new information to add, and the candidate cell state represents potential new content to be integrated. (2) (3) Here, and are weight matrices for the input gate and candidate state, respectively, and and are their corresponding biases. Cell State Update: The new cell state is computed by combining the previous cell state and the candidate cell state, modulated by the forget and input gates. (4) Output Gate and Hidden State: Finally, the output gate decides what part of the cell state should be output. The hidden state is then computed as: ) (5) = ) (6) where and are the weight matrix and bias vector for the output gate. LightGBM LightGBM’s mathematical foundation is rooted in the gradient boosting decision tree (GBDT) framework, and it leverages several key innovations to enhance both efficiency and accuracy: Gradient Boosting and Second-Order Approximation: LightGBM minimizes an overall objective function that combines a loss function with a regularization term. At each iteration, it fits a new decision tree to the negative gradients (first-order derivatives) of the loss function. To improve convergence, it uses a second-order Taylor expansion: Where and are the gradient and Hessian for instance iii, and is a regularization term for the tree. Optimal Split Gain Calculation: When constructing each decision tree, LightGBM determines the best splits by maximizing the gain. For a split that partitions the data into left and right children, the gain is computed as: (8) Here, and (and , ) are the sums of gradients and Hessians for the left (and right) child nodes, while λ and γ are regularization parameters controlling overfitting. Journal of Theoretical and Applied Information Technology 31st May 2025. Vol.103. No.10 © Little Lion Scientific ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195 4420 Histogram-Based Learning: To speed up computation and reduce memory usage, continuous feature values are bucketed into discrete bins (histograms). This approximation allows LightGBM to quickly compute split gains and perform efficient tree learning without scanning all individual data points. Advanced Sampling and Feature Bundling: LightGBM incorporates Gradient-based One-Side Sampling (GOSS) to retain instances with large gradients (which are more informative) while subsampling those with small gradients. It also uses Exclusive Feature Bundling (EFB) to combine mutually exclusive features, reducing dimensionality without significant information loss. Ensemble Stacking Regressor Below are several key equations that mathematically describe an ensemble stacking hybrid model. Assume that we have an input feature vector and a target variable . The stacking ensemble uses multiple base models to generate predictions, which are then combined by a meta-learner. Base Model Predictions: Suppose we have base models denoted as . Each model produces a prediction for the input : Meta-Learner (Stacking) A meta-learner is then trained to map the metafeature vector to a final prediction where is the intercept and are the weights learned for each base model’s prediction. Loss Minimization: Meta-learner parameters are preserved by minimizing the appropriate loss function through the training set. For a typical regression scenario using Mean Squared Error (MSE), the objective is: For quantile regression, where the goal is to estimate a particular quantile (e.g., the median when q=0.5q=0.5q=0.5), the pinball loss function is used: 4. Result Analysis Journal of Theoretical and Applied Information Technology 31st May 2025. Vol.103. No.10 © Little Lion Scientific ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195 4421 Figure 3. Staked Ensemble Quantile Regression Prediction ON Gsbd STOCK Data The graphics show the performance of the stacked ensemble quantile regression model when predicting GSBD stock prices and comparing actual values (black) with predicted values (red). This model accurately follows the actual trends that exhibit strong predictive capabilities, especially in areas with high variance. Although well documented, occasional underestimation and overestimation indicate potential improvements through further optimization. Overall, the ensemble model exhibits robustness of share prognosis with space for fine tuning to improve accuracy in warm areas. Figure 4. True VS Ensemble Prediction Error Bands The graphics show a comparison between true stock prices (black) and ensemble model (red) predictions. The close direction of the ensemble prediction using actual values indicates strong prediction accuracy. The inclusion of fault zones provides insight into model uncertainty, and slight deviations are observed in minor areas. Although the model appears to have a good grasp of variability, occasional underestimation and overestimation highlight areas for possible improvement. Overall, the ensemble model exhibits robust predictive performance with controlled uncertainty. Journal of Theoretical and Applied Information Technology 31st May 2025. Vol.103. No.10 © Little Lion Scientific ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195 4422 Figure 5. Joint Plot: True Vs. Ensemble Predictions The general diagram shows the relationship between true and shared final prices for ensemble plans, showing strong linear correlations. The scatterplot shows that most points are near the ID line, indicating an accurate prediction with minimal deviation. The top and right histograms provide insight into the distribution of actual and predicted values and show similar spreads. Although slight variance is displayed at some points, overall the ensemble model shows high prediction accuracy and minimal distortion. Figure 6. Rolling Mean Of Residuals