Eashan Mathur

This investigation explores if the use of ARIMA modelling for the prediction of stock prices is accurate for the stock prediction of three companies. AutoRegressive-Integrated-Moving-Average (ARIMA), is a function within the R language framework that specializes in prediction based on history.

To explore this topic, the paper starts with introducing Machine Learning alongside with what stocks are and their role in our world. Subsequently, an experiment was carried out which involved the… Show more

This investigation explores if the use of ARIMA modelling for the prediction of stock prices is accurate for the stock prediction of three companies. AutoRegressive-Integrated-Moving-Average (ARIMA), is a function within the R language framework that specializes in prediction based on history.

To explore this topic, the paper starts with introducing Machine Learning alongside with what stocks are and their role in our world. Subsequently, an experiment was carried out which involved the development of an algorithm in R, accepting a company parameter to gather their stock data. With that data, the algorithm split the points into lags and used ARIMA modelling, using the past trends, in order to predict possible future values. To make the stock values work for the model, it was required to convert the regular graphs into difference graphs, depicting the variance in the prices. The model works best when the differenced data is stationary, when it has a mean of 0 and is centered around one value. To receive the most accurate results, the values were computed through log and square root transformations with the logarithmic being the most stationary. The model was then run on three companies: Apple, Microsoft, and Amazon. Due to the future being impossible to fact-check, the model was instead run to predict past values.

After analyzing the model’s results, it was able to predict close values but despite the variance hovering between 1-4% from the real prices, even 2% can be large for companies like Amazon with their prices in the thousands. The effort put into this experiment helped understand how algorithms and models work, but in reality, to make this effective for real world use, the model needs a lot more training and optimization before it can be used for true prediction.
Show less