Time-Series-Analysis-of-Stocks-in-R

#Introduction

We were needed to use R to implement the time­series forecast of stocks in NASDAQ of
which the data was provided. The execution of R code has to be done in CCR. The
approaches used are: ● Linear Regression Model ● Holt­Winters Model ● ARIMA model

#Package Usages

● Forecast​ : Provides the necessary functionality of converting and evaluating the data
for time series analysis. The models that are implemented in the code use the library
function provided by this package ● Fpp​ : This library has functionality that builds over the Forecast package. ● Plotly​ : This is the Graphics library used.

#Evaluation Metric

The MAE (Mean Absolute Error) is a common measure to evaluate error in time series
analysis.

#Goal of the Project

● Based on the above metric we needed to find stocks with best­forecasted
performance, i.e. stocks with the minimum value of Sum MAE value. Stock which are
best fitting the model learned using the Models mentioned above. ● Also we were required to compare three techniques, namely, Linear Regression
Model, Holt­Winters Model, and ARIMA model.

#Implementation and Result of Linear Regression Model

Function used for calculation:

● tslm(trainData ~ trend) ○ tslm = Name of the function ○ trainData = Adjusted close price time series data fetched from the csv file for
the first 744 days. ○ trend = Seasonality used in the formula. Uses “dummy variables” for seasons.
In R, tslm automatically generates seasonal dummies for a ts object

#Implementation and Result of Holt­Winters Model

Function used for calculation:

● HoltWinters(trainData, gamma = FALSE) ○ HoltWinters = An object of class "HoltWinters" ○ trainData = Adjusted close price time series data fetched from the csv file for
the first 744 days. ○ gamma = This parameter will do the exponential smoothing

#Implementation and Result of ARIMA Model

Function used for calculation:

● auto.arima(trainData) ○ auto.arima = Fit best ARIMA model to univariate time series (from
documentation) ○ trainData = Adjusted close price time series data fetched from the csv file for
the first 744 days.

#Problems Faced

● While executing the code, the values were changing as the parameter of various mode
functions were changing. Hence the experimentation process followed was focussed
on getting the actual trends with the minimum error. ● At the CCR, the Single Node 8 core machine was selected in the SLURM script but
this was leading to printing the value 8 times, each time for each core. ● Also, care needed when dealing with the non existing files or files which are empty
since the execution may get halted in between and will be failed job.

#Comparison Between the Models

Different trends were observed for different models. The reason could be different fitting of the
training data. Since the models are trying to learn a generalized representation of data which
results in a Mathematical equation (Linear or Nonlinear, depends upon the model and
parameters used). When the test data is given to this equation, it will generate the values and
we will calculate the value. Now the deviation of this calculated value with respect to the
actual values is the Error. Since we are getting dissimilar trends for different models we can
rightly say that the Equation differs for each case.