Project 2 - Forecasting
Collect data on any topic of interest to you, preferably
something related to your work (application of the techniques from this course
to your work will add value to your organization and to you.) You may choose to
do either Time Series Forecasting or Causal Forecasting.
Time Series Forecasting
Pick a variable you want to forecast, and collect data on its
values in the past (at least 25 time periods, preferably closer to 50). The data
should NOT be annual, since that will prevent the study of seasonality, if any.
Choose monthly or quarterly or daily or hourly data.
Create a scatter plot showing the data over time. Discuss what
approach is most appropriate for forecasting.
Regardless of the above, forecast using all the methods
discussed in class – Naïve, Moving Average, Exponential Smoothing, Regression,
and Classical Decomposition.
Compare all methods using appropriate evaluation criteria
(Bias, MAD, MAPE, MSE, Standard Error)
Causal Forecasting
You must have at least 5 independent variables (can be a mix
of categorical and numeric) and a dependent variable (numeric). The number of
observations will depend on the circumstances, but in general, the more the
better – rule of thumb is to get at least 10 times as many observations as
variables (so for 5 variables, you need 50 observations).
For each variable, show distribution of observations with
frequency charts, mean and standard deviation computations, etc.
Show relationships of each independent variable individually
with the dependent using scatter plots.
Perform regression analysis to show overall model for
predicting the value of the dependent. Remember to eliminate variables that are
not significant, and run regression analysis multiple times until your model has
only significant variables, or until you conclude that nothing is significant.
Interpret
the results and write a report. The report must first briefly describe the
background, what you are trying to predict, what the variables are, and how you
collected the data, before showing the analyses and results. The report must
stand alone – one should be able to understand the salient points of everything
you did without having to look at your spreadsheet.
Report Format
Introduction: What motivates this study? Who
is it important to? Provide general background
Data: How much data was collected? Number of
Observations, the variable(s), the way the variables were
measured, the source of the data
Preliminary Analysis: Scatter Plot(s), and
interpretation of the plots. For time series, what does the plot
tell you about the relationships and the method of forecasting
that might work best? For causal forecasting, draw multiple
plots, and interpret each. How does each X seem to relate to Y?
Forecasting: Forecast Y using all the methods.
Draw graphs of actual vs. forecasted value.
Evaluation: Compute Errors, compute Bias and at least one of MAD, MAPE, MSE (SE). Compare the values across methods.
Conclusion: What is the forecast for the next
time period using the best of the methods for this data?