/Monthly_revenue_time_series_A_liquor_company

Primary LanguageJupyter Notebook

Monthly revenue time series of a liquor company

Introduction

In this project, I consider a time series of monthly revenue of Jim Beam Brands, a liquor company in Iowa State, USA. First of all, I explore the important characterizations of this time series, such as decompositions (both multiplicative and additive), stationary property, seasonal plot ... In the next step, I make some forecasts based on two models: SARIMA and Prophet. Trend changes detection, cross validation and more importantly, the hyperparameters tuning are also included.

Dataset

The dataset used in this project is a part deduced from a huge dataset at Iowa’s state-hosted open data portal, which has around 24 million rows and 24 columns.

For the purpose of this project, I filtered the vendor named Jim Beam Brands, resulting in 2180745 rows. The following columns are considered:

  • Date: date that bottles were sold to retailers.
  • Item Description: names of the sold bottles.
  • Bottle Volume: volume of each bottle.
  • State Bottle Cost: cost the Iowa State pays to Jim Beam Brands for sold bottles.
  • State Bottle Retail: money the State receives from selling bottles to retailers.
  • Bottle Sold: number of bottles sold to retailers.

Methodologies

  • After cleansing the data, I created a time series of monthly revenue of Jim Beam Brands and made some visualizations:

A first glimpse of seasonality as well as the distribution of revenue are also explored.

  • For better understanding, I check the stationary property, seasonality and make decompositions in both multiplicative and additive cases.
  • Finally, the forecasts are carried out based on two models: SARIMA and Prophet.

In SARIMA, I use auto_arima to search for best possible parameters p, q, d and test the result on a training set. Due to fact that this dataset has unexpected trend changes and outliers, two training sets are used to validate the accuracy.

With Prophet model, a cross-validation is implemented, and later used in hyperparameters tuning, including changepoint_prior_scale and seasonality_prior_scale. These parameters are then put into the forecast to effectively the RMSE error.

Results

1. The seasonality can be somewhat suggested by seasonal plot

and the box plots:

2. Decomposition is done by seasonal_decompose from statsmodels.tsa.seasonal

3. Forecasts with SARIMA model

Forecast on a training set containing 80% of our dataset:

The errors in this case are:

On another training set containing info of outliers, the result is better

with better errors

4. Forecasts with Prophet model

The first forecast is done without hyperparameters indicated:

We get an improvement of reducing RMSE error after hyperparameters tuning:

The RMSE error drops from 178302.105 to 150059.697.