Please give me a response to each discussion post telling me what you liked or disliked and provide any other alternatives to each. Provide me a half page on each discussion response.
Discussion #1:
In our week one discussion I highlighted my roles and responsibilities as an eBusiness manager at 21st Century Insurance. As a direct marketer, we are heavily dependent upon the Internet and other technological channels to support our business model and corresponding operating goals. Moreover, I mentioned why forecasting was an essential component, not only on the business side but also in our IT department. eBusiness as a sector of the marketing department must forecast website & mobile device traffic accurately, based upon our on projected spend but we also must ensure that we communicate and compare models with the IT department. This is needed to ensure that they purchase the necessary infrastructure (server space, Akamai accelerator, etc...) in order to realize our forecasts.
Based upon these attributes mentioned previously I feel that a time series model would be more appropriate than a casual forecasting model. With this being said, in our infancy (before web & mobile traffic over took that of our call center) I would argue that we followed a casual model because of a lack of accrued historical data and the informal approach / lack of budget directed towards this emerging channel. Because of this, I feel that we as a business missed out on any chance of a competitive advantage as a result of speed to market and now must play catch up to those giants that realized the importance of this new channel.
In the time series approach I would expect the components of seasonality and linear trend to be the most dominant. For instance, we see noticeable downward trends during off times, weekends, and holidays. Correspondingly, we see upward trends during the last & first days of the month then during seasonal periods (July - Aug. and Jan. - March). Since most insurance policies are issued on six month policy terms one can easily see how these intervals affect or business twice year. The measure of forecast accuracy I would use to evaluate our model would be MAD since the market is ever changing but not at a significance level that we need to be concerned with at an absolute number. For instance, SEO values can vary on a daily basis due to shifting algorithms and due to geo-targeting SEO values differ by location. With this being said, a variance is expected and one should not over analyze fluctuations over small periods of time, but needs to be compared over longer durations to ensure positive linear trends or to account for negative trends.
Discussion #2:
The example that I created last week was forecasting customer demand for foreign sheets of commercial grade printing paper. As I understand the two forecasting models presented in the question, a time series method, a quantitative approach, bases its forecast solely upon historical patterns in the data, while the casual method, a qualitative approach, relies upon the judgment of experts in creating the forecast.
In creating a forecast for foreign sheets of commercial grade printing, it appears as if the casual model would prove to be more useful than the time series model. Although there may be some seasonality in the business - back to school, holiday greeting cards, etc. - there are other variables that come into play that would be less detectable by historical patterns and more readily apparent to an expert. For example, during an election year the demand for commercial grade sheets of paper experiences a large spike. A majority of the work that is done for local and regional elections is done on a very tight budget. Because the foreign sheet is often much less expensive than the domestic sheet, campaign officials choose to go with the cheaper sheet. Another scenario that can elicit an increase in demand is the release of a new pharmaceutical. Delivered along with the latest wonder drug are reams and reams of studies, pamphlets and other propaganda. Both of these variables would be very difficult to detect using a time series approach, but would be readily apparent to a sales professional who is continually communicating with the commercial printer. The difficult part is getting that information to the procurement team while it is still relevant.
The measurement that I would choose to evaluate the model would be the root mean squared error (RMSE). There are two reasons for choosing this measurement. The first reason lies in the power of squaring the errors, which amplifies large errors. Because of the resources that are necessary to import sheets of commercial grade paper are both large - capital tied up for months prior to the arrival of the product, the hours of labor necessary to unload overseas containers, the warehouse space necessary to hold the inventory until it is needed - and often in short supply. Large errors can be crippling in an industry as competitive as the paper industry.
The second reason, as the text explains, is the commonality of the units being forecast and the units being measured. Using the same base units allows for an easier interpretation of the measurement and highlights the magnitude of the errors.