Article Preview
Top1. Introduction
As an integral component of a country's economy, the financial market not only reflects the nation's competitiveness but also carries significant responsibilities in the context of the country's socioeconomic mission. With the rapid development of the socioeconomic landscape, the complexity and diversity of financial markets have been on the rise, leading to the accumulation of vast volumes of financial data. This has also raised higher demands for financial information, making the efficient extraction, analysis, and prediction of financial data a pressing challenge in both academia and industry. Therefore, research into intelligent monitoring and early warning models for financial risks holds substantial practical value. The financial sector generates a plethora of structured and unstructured data, including market trading data, news reports, economic indicators, company financial reports, among other information sources. These data not only come in massive quantities but also typically exhibit highly dynamic and diverse characteristics, reflecting the intricacies and uncertainties of financial markets. Traditional time series analysis methods find widespread application in the field of finance, including autoregressive models (AR) (Kaur, Parmar & Singh, 2023), moving average models (MA) (Xu et al., 2023), autoregressive Moving Average Models (ARMA) (Rapoo, Chanza & Motlhwe, 2023), and autoregressive integrated moving average models (ARIMA) (Wang et al., 2023a). Autoregressive (AR) models are advantageous for their simplicity, intuitiveness, and ease of understanding and implementation. They effectively capture the local patterns and trends in data, offering flexibility by adjusting the order to control model complexity. However, AR models, based on the assumption of linear relationships, may struggle to capture nonlinear dynamics and complex relationships. They are sensitive to initial values, require data stationarity, and may have limited effectiveness when dealing with non-stationary or complex data. Moving Average (MA) models, on the other hand, excel at adapting to short-term fluctuations in data. By considering the moving average of past observations, they reduce the impact of noise and random fluctuations, resulting in a smoother and more stable model. MA models are particularly effective in handling seasonal and periodic time series data. However, they have limitations in modeling trends and long-term dependencies, as they primarily focus on short-term average effects and may not fully capture long-term trends in time series. Additionally, MA models may perform poorly with long-term memory in noise, requiring a careful balance and selection based on the data's characteristics in practical applications. Autoregressive Moving Average (ARMA) models combine the strengths of both AR and MA components. They capture long-term dependencies and trends (via the AR part) while effectively handling short-term fluctuations and noise (via the MA part). ARMA model parameter estimation is relatively intuitive, exhibiting strong adaptability to time series data of different natures. However, ARMA models have limited capabilities in modeling nonlinearity and non-stationarity, requiring prior assurance of data stationarity. Additionally, parameter selection for the model may demand empirical and domain knowledge. Careful consideration is necessary when balancing model complexity and fitting performance, especially when dealing with high-order models to avoid overfitting. Autoregressive Integrated Moving Average (ARIMA) models are widely used, decomposing time series data into trend, seasonality, and residual components for predicting future trends. Although traditional methods perform well in certain situations, they often rely on strong domain knowledge and manual feature engineering. Their ability to handle nonlinear and non-stationary data is limited, and they typically depend on statistical models and rule-based systems, posing constraints when dealing with large-scale, multimodal, and high-dimensional data.