Time series dataset used to predict van rental sales demand.
A National Van Rental company has expenses that are variables based on demand. The largest and most volatile expense is the fleet cost. If the fleet is too large, then massive quantities of car notes are paid on unused vans. If the fleet is too small, then revenues and market shares are lossed.
- What are the peak Days? Weeks? Months? Quarters? Season?
- What are the slow periods? How long do they last?
- How often is the company under fleeted?
- How often is the company over fleeted?
- Can the company optimize fleeting patterns?
- Build an accurate model to predict van rental sales demand.
- Help the Van Rental company optimize variable expenses.
- Macbook: CPU 4; 16gb Ram; With GPU
- AWS EC2: CPU 16; 64gb Ram; No GPU
Raw Data: CSV
- DataFrame: (Rows, Columns)
- 2019 Revenue: (106494, 171)
- 2018 Revenue: (106563, 165)
- 2017 Revenue: (96909, 159)
Prepped Data:
- Total Days: 1,095
- Total Sales Count: 305,353
- The bell curve shape histogram indicates sales follows a Normal Distribution.
- In the boxplot we see the monthly seasonality via the wavy shape.
- The Sales By Month bar chart shows peak months are in the summer.
- The Sales by Day of Week bar chart shows the peak day is Friday.
- Custom Trailing Moving Average Function (TMA)
- ARIMA
- fbprophet
- LSTM
- The lower the error the better the model.
- The Dickey Fuller test is a mathematical test for stationarity.
- This data is NOT stationary.
- In order to use certain algorithms the data must be stationary.
- The Graphs above show RMSE on the y-axis and fbprophet trend hyperparameters on the x-axis.
- The smallest RMSE is indicated by the dotted redline per graph.
- I used 70% of the data to train the model and 30% of the data to test and evaluate model performance.
- The blue line represents the PREDICTIONS and the orange line represents the ACTUALS from the test data.
- The blue PREDICTION line and the orange ACTUALS line OVERLAP nicely, indicating the model is performing well.
Experiment Reality:
- Cars can be purchased and sold every day with no lag.
- There is always an immediate buyer when vans are being sold.
- There is always immediate funding when vans are being purchased.
Experiment Rules:
First Rule. Optimal fleet availability is double the predicted demand.
Second Rule. Any time fleet is greater than one BUT less than optimal, the company will benefit 400 Dollars per car.
Third Rule. Any time fleet is over the optimal mark it will cost the company 400 Dollars per vehicle.
Fourth Rule. Anytime fleet is in the negative it will cost the company 400 Dollars in revenue per vehicle.
Model Application Results
- The highest amount of over fleeting cost happens late December to early January.
- Performance between man and model oscillated between profit and loss.
- If the model was deployed the company would have save $360K thousand dollars for the Los Angeles Location!
- I created a Tableau dashboard with an interactive map showing sales demand by state & by city
- The pie chart shows the split between repeat customers and non-repeat customers.
- The horizontal bar graph shows demand by van type.
- A dashboard like this could be utilized by a company's non technical employees to observe the results of some of these very technical time series analyses.
- Peak Day: Friday
- Peak Months: June & July
- Slow Day: Sunday
- Slow Months: January & December
- The company should look into drastically reducing fleet late December to early January.
For this van rental company COMPLETE fleet automation may be too ambitious, BUT AUGMENTING fleet decisions is very possible and it would contribute immense value to this organization.