Problem Description
Ford Motor Company deals with a product portfolio that consists of three subcategories namely
cars, light trucks and heavy trucks. The lead time for manufacturing planning of any subcategory
is 12 months. Hence they need a forecast of the total auto sales in the US market for the next 12
months on a monthly basis. Some examples of vehicles in each subcategory are shown in the
appendix. Both domestic sales and exports are to be forecasted.
This project only deals with generation of the forecasts and not their application. However, Ford
company management may use the forecasts for the following applications:
● Pre-order various auto parts to make sure production process are streamlined.
● Make proper investment in production facilities since setting up new equipment requires time
and needs to be planned at least 6-12 months ahead of the requirement.
● Synchronize growth strategy to be in line with the sales trend across different auto categories.
● To efficiently distribute the final products between export and import markets.
The forecast period is next 12 months. The following monthly time series have been forecasted
in this project: Car sales, Light truck sales, Heavy truck sales, Car export sales, Light truck
export sales

Description of the data
Data for our analysis has been sourced from the Bureau of Economic Analysis (BEA). This
agency consolidates the data for the entire US market using company resources and industry
reports such as Ward’s automotive monthly report, US passenger car production report etc.
The original data source contains monthly sales figures for the entire united states of 3
subcategories of vehicles. The period is from 1973 to 2016. Ideally we would have wanted the
data for only Ford vehicles but it is not available. We have assumed that Ford management will
adjust the forecasts taking into account their market share for each of the vehicle subcategories.

Forecasting Methodology:
1.Seasonal Naive to establish benchmark: The data exhibits a seasonality of 12 months. Thus,
seasonal naive forecasting includes forecasting monthly sales using the 12 month prior values
in the same time series. This is considered as a benchmark and subsequent methods are aimed
at improving upon this forecast accuracy.
2.Holts-Winter method as the data exhibited trend and seasonality: Since all data series
consistently exhibit trend as well as seasonality we have chosen to first try out Holt-Winters
method. The training and validation period for all-time series are different, they have been
adjusted to achieve the most accurate forecast.
3.MLR with monthly dummy variables: Next we have used multiple linear regression using
dummy variables for each month. Depending on the characteristics of the series both additive
and multiplicative regression methods have been used.
4.Ensemble combinations improve results: Finally, we have tried combinations of Naive, HoltsWinters
and Multiple Linear Regression to create ensembles that can further improve the
accuracy of the forecasts.
5.MAPE and residual graphs used to judge and select a final model: Once these methods are
applied, their validation period residuals are plotted together to gauge the accuracy of each of
them. The final model for a particular data series is selected taking into account the residuals
plot and the MAPE values.

Forecast Risks
1) Economic and Geopolitical Shock: This model cannot predict sudden and unforeseen shocks.
Economic shocks include recessions, market crashes, imposition of heavy import duties,
strict environmental protection laws and spikes in oil prices. Geopolitical shock would
include political turmoil, wars, large scale terrorist attacks and natural disasters.
2) Forecasting sales of individual models: These forecasts only provide the total sales in the US
market. Ford management must adjust these using their market share estimates to get the
actual model wise sales of ford vehicles. There is a risk of uncertainty in the market share
estimates that could make model wise forecasts completely inaccurate.
Conclusions & Recommendations
We forecasted the next 12 months sales for the 3 subcategories with reasonable accuracy and
confidence intervals (See charts in appendix) and performance metrics in below tables.
We have not been able to create a function to quantify the penalty of under forecasting due to
lack of data. But given the data of average prices in each segment and assuming that all lost sales
go to competitors we can create a loss function which can provide the financial impact for the
errors in forecast.
It is recommended that Ford management take into account the forecast confidence intervals into
any production planning exercise. This is because the forecasts have an uncertainty associated
with them. It is further recommended that these forecast models be retrained with each new
monthly data point as and when they are available. Hence rolling forecast should be followed.