|
|
The use of artificial intelligence methods for time series prediction
Tripathi, Ankit
Financial market analysis and prediction have been topics of interest to traders and investors for decades. This thesis presents a comprehensive study on time series forecasting in the dynamic financial market of India, utilising a decade of historical data from Reliance Company's stock prices. The research encompasses three key components: bibliometric analysis for the domain globally, comparative evaluation of time series prediction methods in Indian markets, and implementation of a pre-processing approach incorporating economic factors on the selected models. Every section builds upon the collected information in the preceding section. The bibliometric analysis was used to establish an understanding of prevailing trends in time series forecasting techniques and answer relevant questions in the context of Indian markets to narrow down the scope of the study. This has been done by analysing 2202 documents ranging from the period of 1994-2023 consisting of articles, book chapter, review, book, note and letters in the English language only. The results help in the formation of a different perspective while understanding the overall intellectual landscape of the domain with subsections focusing on field leaders, author's productivity, uprise in domain based on publications and citations, the underlying pattern behind shifts in research areas based on authors keywords and publications that have impacted the domain significantly. The analysis extends beyond academic literature to include patents, providing a real-life state-of-the-art perspective. The results from bibliometrics have been used to select models for comparative analysis. The analysis assesses the performance of diverse time series prediction methods like deep learning algorithms (Long short-term memory model (LSTM)), traditional statistical models (Auto Regressive Integrated Moving Approach (ARIMA)), and advanced ensemble learning algorithms (XGBoost and FB-Prophet) using real-world data from the Indian financial market. The stock prices of Reliance Company serve as a case study, enabling a thorough evaluation of predictive accuracy and errors of the models. Simultaneously, a pre-processing approach has been proposed and implemented, integrating significant economic factors (Gold Price, USD to INR conversion, Consumer Price Index, Indian 10-year yield bond, and Wholesale Price Index) and evaluated with technical metrics (Mean squared error, Mean Absolute Error, R2 Score). The study investigates how the inclusion of these factors impacts prediction accuracy across the selected time series prediction methods. The comparative evaluation of models before and after the pre-processing method sheds light on the evolving predictive accuracy of LSTM, ARIMA, FB-Prophet, and XGBoost. This analysis provides valuable insights into the influence of economic factors on each method's performance. The study showed that the SARIMAX (extension of ARIMA with seasonality and exogenous factors) and XGBOOST performed relatively well with the proposed approach while LSTM with 80% training and FB prophet did not perform as expected in Indian financial markets. This research contributes to advancing the understanding of time series forecasting in the financial market of India, offering practical insights for decision-makers and researchers.
|
guest ::
RSS feed.