This repository contains implementations of various demand forecasting models that can be used for an E-Commerce environment. The models used are ARIMA time series forecasting and Regression using Digital Marketing KPIS. The implementation is done in R.
The file for ARIMA time series forecasting is 'ProductA_timeseries.R'. The implementation uses the following packages:
- 'tidyverse' for data manipulation
- 'forecast' for time series analysis and modeling
- 'evobiR' for sliding window calculation
- 'tseries' for time series analysis
- 'urca' for unit root testing
- 'TSstudio' for residual diagnostics
- 'Metrics' for calculating evaluation metrics such as Mean Absolute Percentage Error (MAPE), Root Mean Squared Error (RMSE) etc.
The sales data of Product A is stored in the file ProductA.csv. The dataset contains the sales quantity of Product A. The data can be extracted from the Google Analytics dashboard of the corresponding E-Commerce website.
- The dataset is split into training and testing sets with a ratio of 80:20.
- A time series object is created from the training data set.
- The training data set is cleaned using the sliding window method.
- The ACF and PACF of the time series object are plotted to check for stationarity.
- The time series object is differenced if needed to make it stationary.
- Tests for stationarity (ADF, PP, and KPSS tests) are performed on the time series object and the differenced time series object.
- The time series object is decomposed to understand its trend, seasonality, and residuals.
- An ARIMA model is fit on the training data set.
- The model is used to forecast the sales for the testing data set.
- The mean absolute percentage error (MAPE) between the actual sales and the forecasted sales is calculated.
- A line plot is created to compare the actual and forecasted sales.
The results of the time series analysis can be found in the console after running the code. This includes the summary of the fitted model, evaluation metric Mean ABsolute Percentage Error (MAPE), forecast values and forecast plot.
The file for multivariate regression is 'ProductA_Multivariate_Regression.R'. This is another model that performs a multivariate regression of sales of "ProductA" on two predictor variables: google clicks and facebook impressions. These are two Digital Marketing KPIs that influence the sales of a product. Google clicks refers to the number of clicks associated with a particular Google Ads campaign that drove the sales of a particular product (here it is Product A). Similarly, Facebook Impressions refers to the number of impressions associated with a particular Facebook Ads campaign that drove the sales of a particular product. The implementation uses the following packages:
- 'tidyverse'
- 'evobiR'
- 'MLmetrics'
- 'Metrics'
The code requires three data sets:
"ProductA.csv" which contains sales quantity data. "ProductA_google_clicks.csv" which contains clicks data. "ProductA_fb_impressions.csv" which contains impressions data.
The code first loads the required packages and reads in the three data sets. Then, the data is cleaned using moving average and normalizing the predictor variables. The data sets are merged into a single data frame and split into training and testing sets.
A multivariate regression model is fit using the "lm" function in R, with sales as the response variable and google clicks and facebook impressions as the predictor variables. The model summary is displayed and the model is used to make predictions on the testing data.
Finally, three evaluation metrics (Mean Absolute Percentage Error (MAPE), Mean Squared Error (MSE), and Root Mean Squared Error (RMSE)) are calculated to assess the performance of the model.
The file for multivariate regression is 'ProductA_Dynamic_Regression.R'. This is another time series model that uses dynamic regression to study the sales of a product and the relationship with two predictor variables (Google clicks and Facebook impressions).
The following libraries are used in the project:
- 'tidyverse'
- 'forecast'
- 'evobiR'
- 'tseries'
- 'urca'
- 'TSstudio'
The sales data, Google clicks data, and Facebook impressions data are read from respective CSV files. A sliding window is applied to calculate the mean of each variable over a 3-time step window, and the resulting series is split into a training set (80% of the data) and a test set (20% of the data). Time series objects are created for each variable using the ts() function from the base library. The sales time series and both predictor time series are tested for stationarity using the kpss.test() function from the tseries library.
The auto.arima() function from the forecast library is used to fit an ARIMA model to the sales time series with the two predictor variables as external regressors. The seasonal argument is set to TRUE to include a seasonal component in the model. The fitted model is stored in the object fit_productA_sales.
The summary() function is used to summarize the fitted model and the checkresiduals() function from the TSstudio library is used to check the residuals of the model.
The results of the time series analysis can be found in the console after running the code. This includes the summary of the fitted model, residual diagnostics, and any other relevant information.
If you have any suggestions or improvements, feel free to create a pull request.