Demantra: Analytical Engine - A deeper look - Part1

Through this post, I would like to share some of findings about how Demantra's analytical engine works. The post is based on Demantra's documents created by Oracle and are available in Oracle Metalink for a detailed study. 

 What is Forecasting

Forecasting empowers one to predict the futuristic tendencies of supply chain, influenced by seasonal and other predictable factors. The result of forecast is a projected curve that has been smoothed to show tendencies and de-emphasize the exceptional variations.

Demantra engine is capable of multidimensional forecasting with mixed modeling techniques. It has been designed to be robust enough for handling large scale installations. It is capable of modeling hundreds of thousands of data sets with various data patterns.

Modeling is performed according to forecast tree, in which each node in the tree is defined by an item-location combination in the Enterprise Data model. Engine tries to generate forecast at lowest level of item-location combination, but in situations where due of data insufficiency's , it is unable to generate a good forecast, then it traverses the Forecast tree to higher level for generation of forecast at an aggregated level. Forecast generated at higher level is split down to lowest level, using a splitting mechanism called “proport”.

Batch verses Simulation

Engine can be executed in two modes, Batch or Simulation.

Batch Run:

• Traverses a large forecast tress, described in database. Each node in tress represents a time based data series that is subjected to forecast.

• Performs model calculation on a large subset of data series (each node of forecast tree)

• Writes the processed data series to a new forecast table in database.

• Executes a post-process procedure which updates new forecast table

Simulation Run:

• Same steps as above are performed but on a selective subset of data, as chosen by user for “what-if” kind of simulation run.

  
  

  Engine Components

  
  

  Engine consists to two components, Engine manager and Engine server. There is one instance of Engine manager and one or more instances of engine server. The engine server scans a portion of the forecast tree according to the traversal rules, and sends the output to the split mechanism. 
  The engine server receives IDs of tasks from Engine manager, which divides the forecast tree into sub-trees (tasks). The engine manager is responsible for the engine run as a whole.  

  
  

  Based on Engine settings done via Engine Administrator application, Engine managers tries to initiate an engine server on list of machines on network having engine registered and server component installed.

  
  

  Reference “Details of Engine Manager & Server”. (source: Oracle Demantra Engine Guide v.5.2.6)

  
  

  Causal Factors

  
  

  Factors which cause the deviation from a trend, are known as Causal Factors. E.g. a sales promotion, due of it’s result you expect sales to rise. On the other hand, a sales promotion by your competitor can impact your sales negatively, which also becomes a causal factor.

  
  

  Causal factors such as promotions, marketing campaigns and production stoppages are essential components when analyzing demand pattern.

  
  

  Global Causal Factors effect all the nodes of forecast tree, seasonal CFs are normally global.

  
  

  Local CFs can be different for each item-location combination.

  
  

  Event is another word for CF, it can be loaded easily with Code, value and intensity information. Event information is not necessarily required to be loaded in future.

  
  

  
  

  ............to be continued.