Optimization of LSTM Algorithm Through Outliers – Application to Financial Time Series Forecasting

Abstract. The information available on the internet nowadays is diverse and moves very quickly. Information is becoming easier to obtain by the general public with the numerous online media outlets, including news portals that provide up-to-date information insights. Various news portals earn revenue from advertising using pay-per-click methods that encourage article writers to use clickbait techniques to attract visitors. However, the negative effects of clickbait include a decrease in journalism quality and the spread of hoaxes. This problem can be prevented by using text classification to classify clickbait in news titles. One method that can be used for text classification is a neural network. Artificial neural networks use algorithms that can independently adjust input coefficient weights. This makes this algorithm highly effective for modeling non-linear statistical data. The artificial neural network algorithm, especially the Long Short-Term Memory (LSTM), has been widely used in various natural language processing fields with satisfying results, including text classification. To improve the performance of the neural network model, adjustments can be made to the model's hyperparameters. Hyperparameters are parameters that cannot be obtained through data and must be defined before the training process. In this research, the Long Short-Term Memory (LSTM) model was used in clickbait classification in news titles. Sixteen neural network models were trained with different hyperparameter configurations for each model. Hyperparameter tuning was carried out using the random search algorithm. The dataset used was the CLICK-ID dataset published by William & Sari, 2020[1], with a total of 15,000 annotated data. The research results show that the developed LSTM model has a validation accuracy of 0.8030, higher than William & Sari's research, and a validation loss of 0.4876. Using this model, researchers were able to classify clickbait in news titles with fairly good accuracy. Purpose: The study was to develop and evaluate a LSTM model with hyperparameter tuning for clickbait classification on news headlines. The thesis also aims to compare the performance of simple LSTM and bidirectional LSTM for this task. Methods: This study uses CLICK-ID dataset and applies different text preprocessing techniques. The dataset later was used to build and train 16 LSTM models with different hyperparameters and evaluates them using validation accuracy and loss. This study uses random search for hyperparameter tuning. Result: The results of the study show that the best model for clickbait classification on news headlines is a bidirectional LSTM model with one layer, 64 units, 0.2 dropout rate, and 0.001 learning rate. This model achieves a validation accuracy of 0.8030 and a validation loss of 0.4876. The results also show that hyperparameter tuning using random search can improve the performance of the LSTM models by avoiding zero probabilities and finding the optimal values for the hyperparameters. Novelty: This study compares and analyzes the different preprocessing methods on text and the different configurations of the models to find the best model for clickbait classification on news headlines. The study also uses hyperparameter tuning to tune the model into the best model and finding the optimal values for the hyperparameters.

An accurate electrical Short-term Load Forecasting (STLF) is an eminent factor in the power generation, electrical load dispatching and energy planning for the power supply companies, specifically in developing countries. This paper proposes a novel temporal feature selection-based Long Short-term Memory (LSTM) model developed by the combination of standard Artificial Neural Network (ANN) layer and LSTM for electrical short term load forecasting. The LSTM model has excellent capability of predicting the stochastic nature of an hour ahead electrical loads. The standard ANN layer consisting 11 neurons is used as an input to LSTM cells. Such a combination of ANN layer with LSTM was never proposed before. The proposed model accommodates variations in weather as well as temporal inputs like humidity, holidays, and date-time features in the hourly load data of the power supply company situated in Johor, Malaysia. This paper gives the insights of hyper parameter tuning to capture the more generalized electrical load patterns in the dataset without compromising the time complexity of the proposed model. The proposed approach was compared with five existing approaches, namely: ANN, LSTM model 1, LSTM model 2, LSTM model 3 and Convolutional Neural Network-LSTM (CNN-LSTM) using hourly load dataset of Johor. The experimental results demonstrate that the proposed approach outperformed the existing approaches in terms of root mean square error, mean absolute percentage error and Diebold-Mariano statistical inference test within 95% confidence interval.

Many countries around the globe, including Tanzania, are investing in renewable energy, partly because of the need to increase electric power generation and meet the ever-increasing electricity demands, and partly because of the need to combat global climate change caused by carbon emissions from fossil fuel-based electricity production. Kishapu Solar Project, located in Kishapu District, Shinyanga Region, Tanzania, is one of the renewable energy projects with a planned generation of 150 MW of electric power once completed, using solar photovoltaic technology. Accurate forecasting of solar radiation is important in managing and maintaining solar energy plants. Inaccurate forecasting of solar radiation can lead to poor harvest of solar energy and poor management of solar energy plants, which directly impacts electricity contribution to the national electricity grid. Long Short-Term Memory (LSTM) and its variant, PeepHole LSTM, are examples of Deep Learning models that can be used to predict solar radiation. The purpose of PeepHole LSTM is to introduce “peephole” connections in the LSTM model, which allow LSTM gates to directly inspect memory cell state and hence have richer information to process. However, there is still a research gap on the impact of “peephole” connections on the prediction performance of the LSTM model, especially in the context of forecasting solar radiation in Kishapu District, where the Kishapu Solar Project is located. The purpose of this study is to fill this research gap by developing both LSTM and PeepHole LSTM models and comparing their performances in forecasting solar radiation in Kishapu District. The 14-year Kishapu District daily solar radiation (Downward Shortwave Radiation) dataset was first pre-processed by scaling (normalizing) it using Standard Scaler and then split into training set (the first 70%), validation set (the next 15%) and test set (the last 15%) before being utilized to train and test the developed LSTM and PeepHole LSTM models. Both LSTM and PeepHole LSTM models used the previous 5 days' solar radiation values as inputs to predict the output (next day (6th day) solar radiation value). The trained and better-performing model between the two was finally saved in .h5 format and integrated with a Gradio based Web Application to provide a user interface for officials in Kishapu District and at Kishapu Solar Project to predict daily solar radiation. The findings from performance evaluation revealed that the LSTM model outperformed the PeepHole LSTM model in predicting solar radiation by achieving test Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE) scores of 26.22 W/m2 and 36.65 W/m2, respectively, compared to MAE and RMSE scores of 30.84 W/m2 and 41.88 W/m2, respectively, achieved by the PeepHole LSTM model, concluding that “peephole” connections reduce the performance of the LSTM model rather than increasing it when used to forecast solar radiation in Kishapu District. This study’s recommendation to Artificial Intelligence (AI) software developers and researchers is to use the LSTM rather than the PeepHole LSTM Deep Learning model when forecasting solar radiation in Kishapu District and in locations with similar climatic conditions

Traffic safety has been thought of as a basic feature of transportation, recent developments in civil aviation have emphasized the need for risk identification and safety prediction. This study aims to increase en-route flight safety through the development of prediction models for flight conflicts. Firstly, flight conflicts time series and traffic parameters are extracted from historical ADS-B data. In the second step, a Long Short-Term Memory (LSTM) model is trained to make a one-step-ahead prediction on the flight conflict time series. The results show that the LSTM model has the greatest prediction effect (MAE 0.3901) with comparison to other models. Based on that, we add traffic parameters (volume, density, velocity) into the LSTM model as new input variables and issue a comprehensive analysis of the relative predictive power of traffic parameters. The accuracy of prediction model is validated with a mean error of less than 3%. Based on the improvements of model performance brought by traffic parameters, LSTM models with a single traffic parameter are proposed for further discussion. The results illustrate that volume is the most important factor in promoting prediction accuracy and density has an advantage of improvement in the aspect of model stability.

Wearable electroencephalography (EEG) devices for seizure detection accuracy and reliability are deep learning (DL) applications in the field of epilepsy diagnosis. In this study, we sought to increase the accuracy of seizure detection using advanced DL algorithms on the Children’s Hospital Boston - Massachusetts Institute of Technology (CHB-MIT) EEG database. First, a fully convolutional network (FCN) was trained and assessed using accuracy and recall/precision metrics, and the early stopping technique was used to avoid overfitting. To assess the performance, the FCN was evaluated in terms of various metrics, including accuracy, precision, recall, F1-score, and receiver operating characteristic (ROC)-area under the curve (AUC). In addition, two-dimensional (2D) convolutional neural networks (CNNs) and long short-term memory (LSTM) models were used to model the database, and their performance was thoroughly measured using different metrics, graphs, and confusion matrices. Using LSTM variants, such as standard LSTM, bidirectional LSTM, stacked LSTM, and LSTM attention mechanisms, hybrid convolutional LSTM (ConvLSTM) models were trained and compared. The comparison was conducted based on the training and validation accuracy and loss, as well as the graphs resulting from the precision–recall curves. Apart from DL approaches, EEG signal analysis using time–frequency techniques, such as wavelet transform and short-time Fourier transform, has also been investigated. These methods assisted in the analysis of the time–frequency features of EEG signals in combination with DL models. This study demonstrates that the performance of wearable EEG devices can be augmented using a combination of DL and seizure signal processing techniques. The FCN achieved an accuracy of 92%, with a recall for seizures of 33%, an F1-score of 0.03, and strong ROC-AUC results. The 2D CNN achieved 96% accuracy, a seizure recall of 70%, an F1-score of 0.12, and an ROC-AUC score of 78%. The baseline LSTM struggled with effectiveness at 53% accuracy with a seizure recall of 18%. In contrast, the LSTM model, which incorporated synthetic minority oversampling technique (SMOTE) balancing, was able to reach improvements of up to 89% accuracy, with a precision of 91%, a recall of 86%, an F1-score of 0.89, and a strong ROC curve performance. Among the models, the LSTM with SMOTE was the best performer, with 89% accuracy, 91% precision, 86% recall, and an F1-score of 0.89. These results provide evidence that applying techniques for data balancing in combination with certain DL network architectures significantly improves the detection of seizures using wearable EEG devices worn on the body. We believe that real-time monitoring and high-performance systems are feasible using optimized DL frameworks. The analysis of the performance of different models allows for the understanding of the possibilities of optimizing the architectures of DL algorithms for the modern diagnosis of epilepsy in real time. The source code used to carry out the experiments is publicly available at CHB-MIT EEG Dataset Python Scripts (https://www.kaggle.com/code/adnankust/adnaneeg1).

The effective integration of photovoltaic (PV) systems into the grid which facilitates better planning and resource allocation depends on accurate day-ahead solar power forecasts. This paper proposes a Long Short-Term Memory (LSTM) model for forecasting solar power generation one day in advance with the algorithm inspired by Ant Colony Optimization (ACO). A year's worth of real-time data for a 4.5 kW PV system is used to train and test the model with input characteristics including temperature, PV power production and solar irradiation. Over the course of three seasons—summer, winter, and monsoon—the performance of the ACO-optimized LSTM model is contrasted with that of a simple LSTM model and an LSTM model enhanced by particle swarm optimization (PSO). This encompasses metrics like R-squared (R²) values, Mean Absolute Error (MAE) and Root Mean Square Error (RMSE). The PSO-optimized LSTM model (MAE of 0.31 kW, RMSE of 0.38 kW and R² of 0.88) and the baseline LSTM model (MAE of 0.35 kW, RMSE of 0.42 kW and R² of 0.85) couldn't match the ACO-optimized LSTM model's summertime MAE of 0.27 kW, RMSE of 0.33 kW and R² of 0.92. Similar trends also surfaced throughout the monsoon and winter seasons, with the ACO-optimized model consistently outperforming the others. In the winter, for instance, it obtained an R² of 0.93, an RMSE of 0.26 kW and an MAE of 0.22 kW. With regard to hyper parameter optimization and overall performance, these findings demonstrate the ACO-optimized LSTM model's superiority over the baseline and PSO-optimized LSTM models highlighting its improved prediction precision for all seasons. The effectiveness of ACO in bolstering solar power forecast models is confirmed by this study and provides a solid basis for deep learning model improvement in practical PV applications.

Effective monitoring of tool wear status can improve productivity and reduce losses. In previous studies, extensive feature selection was required when using the traditional machine learning method. The gating mechanism in the traditional long short-term memory (LSTM) model may incur information loss and a weaker representation of global sequential dependencies in handling long sequences. This paper aims to enhance the performance of the LSTM model in tool wear prediction by combining feature and temporal attention. Firstly, the original vibration signal is divided into sub-sequences and related features extracted. Secondly, the ability to capture global sequential dependencies using the LSTM model is improved by feature and temporal attention. Finally, a fully connected layer is used to predict tool wear values. Compared to traditional LSTM, the proposed method performs best in three evaluation metrics, RMSE, MAE, and the coefficient of determination.

This study aims to analyze user sentiment toward the NU Online application by applying a Long Short-Term Memory (LSTM) model. The research was based on 13,576 user reviews collected from the Google Play Store, which underwent preprocessing, sentiment labeling, and Exploratory Data Analysis (EDA). To ensure balanced classification, undersampling was used, resulting in 6,612 reviews equally divided into positive and negative classes. The text data was processed using tokenization and padding before being input into the LSTM model. Model training involved the use of binary crossentropy, Adam optimizer, EarlyStopping, and ReduceLROnPlateau techniques. The model achieved 93% precision, recall, and F1-score, with low error rates and strong generalization ability. EDA results showed that positive feedback mainly focused on worship features like salat schedules, while negative reviews addressed technical issues such as the azan sound. User review peaks occurred during religious periods and major updates. A Gradio-based web interface was also developed to display results and enable user-friendly access to visual sentiment insights. This implementation proves the practical potential of integrating LSTM with an interactive platform for effective sentiment analysis

This study examines the efficacy of three distinct machine learning models for sentiment analysis: Bidirectional Long Short-Term Memory (Bi-LSTM), Long Short-Term Memory (LSTM), and Logistic Regression. One of the most important tasks in natural language processing is sentiment analysis, which entails categorizing reviews as positive or negative. Fifty thousand sentiment-labeled movie reviews make up the dataset. To determine the most precise and effective technique for sentiment categorization, we put these models into practice and evaluated their respective performances. The highest accuracy of 89.42% is achieved by Logistic Regression, with precision of 88.35%, recall of 91.01%, and F1-score of 89.66%. The LSTM model, well-known for capturing temporal dependencies in text, obtained an accuracy of 86.23%, precision of 89.60%, recall of 82.22%, and an F1-score of 85.75%. The Bi-LSTM model, which processes input sequences in both forward and backward directions to better capture context, demonstrated accuracy of 87.65%, precision of 88.67%, recall of 86.54%, and an F1-score of 87.60%. The results show that although Logistic Regression performs better in accuracy, the Bi-LSTM model offers a substantial trade-off between precision and recall, making it a viable option for sentiment analysis applications. Despite its minor accuracy lag, the LSTM model provides important insights into the sequential relationships of the text data.

Wave runup prediction is necessary for offshore structure designs and early warnings. Data-driven methods based on machine learning have inspired reduced-order solutions for wave–structure interaction problems. This study provides the quantification of deep learning algorithms’ potential for wave runup prediction. Two prominent deep learning models were utilized to predict the wave runups along the fore column of semisubmersible under head seas. The long short-term memory (LSTM) and the temporal convolutional networks (TCNs) were comprehensively compared based on the datasets from a model test carried out in the deep ocean basin. The LSTM and TCN model structures were optimized to compare prediction accuracy and computational complexity reasonably. The results reveal that (1) both developed TCN and LSTM models had a satisfied prediction accuracy of over 90%. Their predictions were extended to 10 s into the future with accuracies over 80% and 45%, respectively. (2) With the noise-extended datasets, the LSTM model was robust to noises, while the TCN model showed better prediction performance on the extreme wave runup events. (3) The incident wave and dominant rotation provided the major information for wave runup prediction. TCN and LSTM models’ prediction accuracies were 91.5% and 89.3% based on the simplified input tensors composed of incident wave and pitch. The comparison showed the great potential of the TCN model to predict the nonlinear wave runup with less time and memory costs. The input tensors’ design and optimization based on physical understanding also play a significant role in the prediction performance.

This study utilizes Long Short-Term Memory (LSTM) networks to predict stock prices in the Nepal Stock Exchange (NEPSE), aiming to support investment decision-making. LSTM’s capability to model= temporal dependencies enables it to outperform traditional models in forecasting accuracy. The research uses historical data from insurance companies listed in NEPSE between 2020 and 2024. Evaluation metrics including MAE, MSE, and RMSE demonstrate the effectiveness of the LSTM model. The results are valuable for investors seeking to minimize risk and optimize returns through data-driven strategies. This research employs a Long Short-Term Memory (LSTM) model to predict stock prices in Nepal Stock Exchange (NEPSE). It aims to evaluate accuracy and effectiveness of LSTM in stock price forecasting. LSTM networks, known for their ability to capture complex temporal dependencies, are applied to analyze historical stock price trends. The findings suggest that LSTM models can significantly improve prediction accuracy and offer valuable insights for investors and financial analysts. This paper outlines the methodologies used, presents detailed results, and discusses the implications for investment decision-making.

High-resolution prediction of the spatial distribution of PM2.5 concentrations in China using a long short-term memory model

Real-time anomaly detection based on long short-Term memory and Gaussian Mixture Model

Tanzania heavily depends on United States Dollar (USD) foreign currency to import various goods and services into the country. Failure to correctly forecast exchange rates between USD and Tanzanian Shillings (TZS) may pose risks such as inability to import intended goods and services, possibility of losing money in stock exchange markets and other investment businesses in case of unexpected currency appreciation or depreciation as well as poor investment decisions in foreign exchange markets. To address this, this study has developed and comparatively evaluated performances of LSTM (Long Short-Term Memory) and BiLSTM (Bidirectional LSTM) deep learning models for forecasting daily USD to TZS exchange rates. The findings reveal that, BiLSTM model outperforms LSTM in forecasting daily USD to TZS exchange rates, achieving a MAPE (Mean Absolute Percentage Error) score of 0.363 on test set (unseen data) compared to a MAPE score of 1.471 achieved by LSTM model. This study recommends to the prospective Artificial Intelligence (AI) researchers and software developers to use BiLSTM instead of LSTM model to forecast (predict) USD to TZS exchange rates. Also, this study has developed USD to TZS exchange rates dataset which can be used by AI researchers, saving them time and costs involved with creating datasets from scratch. This study has also developed ready to use BiLSTM and LSTM models which can be used by Tanzanian business men and women involved in stock exchange markets, foreign exchange markets and other businesses, to predict daily USD to TZS exchange rates and make appropriate business and investment decisions.

This study demonstrates the application of deep learning approach specifically the deep learning for cyberfraud incidence classification and time series prediction in the South African financial institutions. Secondary data from the South African Banking Risk Information Centre (SABRIC) was employed and the data was trained under the deep learning paradigm using the Long Short-Term Memory (LSTM) model and adaptive moment estimation (ADAM) algorithm for fraud incidence classification and time series prediction of fraud incidences. Overall, there were 94.1% correct classifications as opposed to 5.9% incorrect classifications. Moreover, the accuracy, precision, recall and F1-score of the LSTM classification model were 71.668%, 87.5%, 99.1% and 78.78% respectively. This indicates that the developed LSTM model is suitable for classification purposes. In addition, the model’s performance improves as new datasets are fed in. This is evident as the root mean square error (RMSE) reduced from 253.5116 obtained initially to 150.9 after new data was fed in. This study contributes conceptually, theoretically and empirical to knowledge on cyberfraud mitigation. The results show that the LSTM model can be deployed for fraud classification and time series analysis of fraud incidences. The outcome of this study may promote cyber resilience and sustain the fight against the perpetration of cyber-related fraud in South Africa’s financial institutions. The use of the LSTM model for cyberfraud classification and time series prediction of cyberfraud incidences in the South African financial institutions demonstrated in this study is unique.