In planning and decision making, forecasting has long been a key component [1]. Making predictions about upcoming trends or occurrences based on historical data and current knowledge is the process of forecasting. It is employed to direct decision-making and policy-making across a variety of disciplines, including economics, finance, weather, and politics [2]. Many forecasting methods are available, including time series analysis, econometric modelling, and machine learning [3]. The integrity and accessibility of information, the difficulty of the underlying system, and the technique employed to make the forecasts may all have an impact on how accurate forecasting is [4].Despite the inherent uncertainty and limitations of forecasting, it is a valuable tool that can help organizations and individuals make better decisions and plan for the future.
Forecasting is highly valued in contemporary culture. Climate change scenarios affect government and corporate policy [5]. Employment and investment are impacted by economic predictions [6]. The COVID-19 virus was stopped from spreading in 2020 by national lockdowns and border restrictions as a result of predictions regarding its spread [7]. According to [8], accurate machine learning (ML) models capable of predicting outcomes across various domains should promote better decision-making at scale and enhance the overall quality of ML.
In order to forecast future events or trends, news items and content on social media are used as a source of data. This method analyses the emotion and substance of news items using natural language processing (NLP) and machine learning techniques to forecast upcoming events or trends [9]. Online news data may be used to forecast a variety of events, including stock market movements, political outcomes, and natural disasters, and has the benefit of offering real-time information [10]. Yet, it's crucial to remember that the credibility and objectivity of the news sources used might have an impact on the accuracy of forecasts produced using internet news data [11].
According to [12] and [13], there are two fundamental approaches to prediction in the forecasting field: statistical forecasting and judgmental forecasting. Traditional statistical time-series prediction models like autoregression or ML time-series models are used to make predictions in statistical forecasting [14] [15]. The models are built and fine-tuned by humans, but individual projections are not changed. This approach effective when the variable being anticipated has a large number of prior observations and a small distribution shift. In contrast, human forecasters employ their own judgement while making projections in judgmental forecasting. To enhance their predictions, forecasters often utilize statistical models along with various sources of information, such as news dataset, past experience and priori reasoning. When historical data is scarce or susceptible to distribution shift, this makes forecasting possible for problems [16]. ML models may be able to provide certain advantages over human forecasts. Models can understand language or data far more quickly than humans can, and they can recognize trends in complex, high-dimensional stuff that individuals cannot. Because they already know the results, people cannot be trained on historical data in a way that simulates true forecasting when it comes to learning, but historical data may be utilized for ML models.
We provide a new dataset for evaluating the propensity of deep hybrid learning models to forecast events, since our research goal is to work with Pakistan news dataset. The dataset includes news text next to each question and questions regarding Pakistani politics, economy, etc. The questions we gathered are those that Pakistan's average man is most likely to have [17]. Selection of the Questions is based on public interest. A dataset's data is set up to date so that deep learning and machine learning models may learn to remember prior events. This dataset is the only one of its type for anticipating Pakistani news.
We use a deep hybrid learning model on our dataset to demonstrate how language systems can be trained using data from our news dataset on historical forecasting issues. Performance benefits from information retrieval increase with system size. Our forecasting model has a 97% accuracy rate. Our forecasting model demonstrating that our dataset is a difficult challenge for ML and Dl.
Fig.I. Dataset example that contain questions and news text collected against a question. To get results for forecasting, we train a deep hybrid learning model.
1.1. Contributions
The following are the research's key contributions.
1) Created the first comprehensive forecasting news dataset that includes only Pakistan-related news.
2) Focus on news analytics, a subfield of news analysis, for news forecasting.
3) Extensive examination and analysis of the developed dataset's performance utilizing the deep hybrid learning model (a combination of machine and deep learning algorithms).
The present research study is structured into distinct sections. Section II presents the literature review, whereas section III delves into the methodology, covering aspects such as the creation and evaluation of the generated dataset, feature extraction, and the experimental setup of the employed models. Section IV analyzes the results and discusses them, while section V presents the conclusion and proposes future research directions.