This is a distribution channel with a single tier at the dealer/partner (hereafter referred to as "partner" or "dealer" interchangeably) that interacts directly with the customer. The dealer is the keystone in the competitive contention for market presence. The relevant business event to be modeled in this example is the dealer. The dealer captures the direct order-taking customer relationship and represents the ordering point for the client.
New order input is the standard for forecasting dealer performance since the sale occurs when the product is installed, accepted, booked and billed. There are too many other processes involved in measuring sales and the concept being tested here is the dealer performance. The ability to ship and support may be a feature input, but is not the specific target metric for this model. New order input is strictly new orders without ongoing maintenance or service contracts and future spares orders. The monthly and quarterly order input for 16 prior quarters is readily available. This information is updated on a quarterly basis.
This data table displays the features, their rank and their weight in projecting the order input for all dealers taken as a whole. These are the features used in the projections.
This chart is interactive. Scroll vertically to observe all feature variables.
The Demand Index is a standardized measure of the market demand for the client's products based on an arbitrary past year. The index is composed of published market information and trends in historical demand for the client's products.
The Price Index is a standardized measure of the client's historical prices as compared to competitor's prices. The sales and dealer staff responses were combined with historical data in an LSTM analysis for trend.
The client has maintained classification data for each dealer that includes region, training levels, ranking, class, territory, credit ratings, etc. This data identifies classes such as 'northwest territory, premier class, training level 5, credit rating 1, etc.' that indicate dealer ability, committment and regional influences.
The client maintains an active internet site that enables the partner to be in nearly constant contact with the client. Records from that site include customer registrations, quote activity, order closes, engineering inquiries, etc. that serve to identify near-term activity levels for each dealer. This data is updated continuously.
Ratings and scores from the account sales, field engineering and field installation teams have been maintained by the client for the client's assessment of the dealer's capabiities. Experience shows that these ratings are often strong predictors of future dealer performance.
The quarterly responses to satisfaction inquiries have been maintained. This amounts to approximately 20 questions from several employees of each dealer each quarter. The questions being added and removed each quarter leave rows and columns of blanks. Incomplete entries by dealer staff limiting their responses to their area of expertise also leaves bocks of blanks. This combination creates large blocks of missing data that has prevented satisfaction response from being predictive variables in the past. This is the only source for measuring the dealer's scoring on the client's performance. Brightfield has implemented a replacement method for this missing data restoring its usefulnesss.
Whether or not promotions for price or other incentives are effective is often a question for anyone engaged in dealer channel management. The time since a price or product promotion has begun or has ended is likely a predictor of dealer performance. The deep learning techniques incorporate these time spans in its analysis.
Sales quotas are not used as a performance predictor. Sales quotas are too involved in sales management procedures and sales personnel compensation to be objective data.
Both market and price life cycles are included in the analysis. These are represented as indices based on the first quarter levels. Developing these indices is the subject of another analytical process not included here. These input variables are represented in this example as the 'Price_Index' and 'Demand_index'.
Sparse data often renders some of these informaton categories ineffective or unusable, in some analytic techniques, as predictor variables and they are subsequently discarded. We have developed a sparse data technique that improves empty data replacement in the data thereby allowing normally discarded data to be retained as a potential predictor variable.
Brightfield advises avoiding accounting postings in any analysis of a sales forecast because the timing and amount of accounting postings are changed by other uncontrollable events in the accounting procedures. Accounting data is posted at fiscal closings and is adjusted according to accounting standards. Unless the accounting data can be reversed to real time and unadjusted, it is best to use some other data.
Lou Carvalheira at Cisco leads the Cisco Model Factory predicting customer order input using tools from H20.ai. The investment in this project is significant and achieved extraordinary results. He made a presentation that is publicly available.