Tuesday, October 7, 2014

Predicting Monthly Car Sales: The Residuals are the Story

I'll produce predictions for US car sales by manufacture every month. There are already several blogs that describe the industry and sales that do a great job. Autoblog by the Numbers and Counting Cars are some to mention. 

Unlike their analysis, I'll try to focus on the residuals (the stuff I can't predict) to tell the story. To highlight the difference, I think its instructive to look at what Autoblog mentioned.

The Autoblog article (link above) highlights Mitsubishi for increasing sales. However, my prediction for Mitsubishi sales are pretty much exactly what the sales were. In essence, given this model, we didn't learn much. On the other hand, Land Rover and Jaguar had the largest residuals (in percent terms) and Land Rover and Acura had the largest deviance (Residual / Variance). I think these results are more telling because we didn't predict them correctly; something might have changed. 

I'll publish this on a new blog: datAutomotive.

Here is a  Shiny App For Car Sales and below are graphs / tables of my current analysis and future predictions. 







Predicted Values 9/14Actual Values 9/14log(Predicted/Actual)Deviance
Acura488.036576.3330.1661.731
Audi632.252621.542-0.0170.326
BMW1,061.3291,066.0830.0040.032
Buick649.648727.7500.1140.571
Cadillac568.021576.2080.0140.089
Chevrolet6,284.5946,411.3750.0200.135
Chrysler1,058.3921,199.2080.1250.710
Dodge1,892.4651,834.167-0.0310.172
Ford7,374.6697,177.542-0.0270.285
GMC1,602.2601,594.542-0.0050.029
Honda4,434.7804,349.625-0.0190.174
Hyundai2,291.3062,333.7500.0180.271
Infiniti315.217326.5420.0350.234
Jaguar37.83447.5830.2290.908
Jeep2,491.3782,301.292-0.0790.853
Kia1,853.7341,692.833-0.0911.115
Land.Rover159.975129.417-0.2121.816
Lexus1,011.967910.500-0.1061.189
Lincoln285.903302.3750.0560.394
Mazda1,074.272999.167-0.0720.743
Mercedes.Benz1,186.2201,230.1250.0360.364
Mini196.975175.792-0.1140.626
Mitsubishi231.582231.5830.000000.00002
Nissan4,176.9953,963.250-0.0530.516
Porsche156.496150.292-0.0400.408
Subaru1,727.5971,729.8750.0010.026
Toyota6,928.4726,059.458-0.1341.309
Volkswagen1,123.4121,083.167-0.0360.407
Volvo161.531194.4580.1861.070




Predicted Values for 10/14
Acura521.878
Audi590.125
BMW1,051.492
Buick666.165
Cadillac533.468
Chevrolet5,693.145
Chrysler1,097.288
Dodge1,493.426
Ford6,694.012
GMC1,572.525
Honda4,049.582
Hyundai1,998.896
Infiniti262.475
Jaguar40.422
Jeep2,132.652
Kia1,650.236
Land.Rover160.884
Lexus944.946
Lincoln272.964
Mazda894.234
Mercedes.Benz1,184.991
Mini183.743
Mitsubishi193.982
Nissan3,629.955
Porsche158.540
Subaru1,694.661
Toyota5,895.316
Volkswagen923.242
Volvo153.691

2 comments:

  1. Predicting such things as car sales, either in aggregate or by make/model, based on time series data is a failed effort. Car sales, in aggregate, are a function of median income and interest rate. At the make/model level, they're a function of specific events, e.g. infotainment system, GPS, ABS, overall design. Time series tells one nothing about future events which will drive increased sales of specific makes/models.

    ReplyDelete
    Replies
    1. I think you're wrong. You can learn somethings from past observations to learn about the future (foremost: trends and seasonality). And just about all time series analysis takes into account previous values.

      I agree there's a lot the model isn't capturing (and there for sure is room for improvement), but to suggest that time series can't predict ANYTHING is bizarre.

      For example look at the Audi Time Series here: https://sweiss.shinyapps.io/CarSales/.

      This model takes into account seasonal and growth factors that have been fairly stable over time. It seems natural (as a first approximation) to assume these trends will continue.

      Delete