Demantra engine tuning
Executive Summary
Many companies continue to struggle with poor forecast numbers even after implementing Demantra’s Demand Management (DM) or Advanced Forecasting (AFDM) modules. The promise of more accurate forecasting post-Demantra implementation remains unfulfilled as demand planners are forced to resort to manual overrides for a large number of records, leading to a lengthy forecast revision process which causes significant delays in the overall demand management process.
The most common reason for poor forecast numbers generated by Demantra is that the engine is not tuned to account for customer-specific dataset requirements.
Ideally, during implementation, a lot of time should be spent analyzing the dataset and configuring the engine parameters taking into account the customer’s specific data model. Unfortunately, it has been found that the Demantra engine tuning exercise has the lowest priority and is often left until after the start-up period.
Many times during the implementation of Demantra, both consultants and business users are so focused on satisfying the requirements related to worksheets, series, workflows, etc. that they tend to take Demantra’s better forecast accuracy for granted and ignore doing due diligence to tune Demantra’s engine.
Also, since Demantra engine tuning is a specialized skill; it requires a deep understanding of various factors that contribute to better forecast accuracy and an understanding of various engine parameters that need to be configured for best results.
While this is a specialized area and should be done by highly skilled and experienced consultants, Demantra users and demand planners should also be familiar with the different factors that can influence forecast accuracy.
There are several factors that influence the accuracy of the Demantra forecast, but some of the most important are listed below:
• Demand data profiles
• Nodal tuning
• Causes/Promotions
• Forecast tree
• Ratio function
Demand data profiles
The first step to better forecasting for Demantra is to understand the different demand profiles that apply to the client’s business.
The demand data pattern can be intermittent, regular, uniform, etc. and knowledge of these demand patterns for the various products would help configure Demantra to use the appropriate statistical model for forecasting.
Oracle Demantra uses different statistical methods and algorithms to project future demand. The Demantra DM model uses eight statistical methods, while the Demantra AFDM uses fourteen different methods for statistical forecasting. The Demantra DM and AFDM modules use the Bayesian approach to generate the final forecast for a specific combination of item and location.
The Bayesian approach combines the results of individual models. Each model is evaluated, and each model, in turn, tests a series of user- and system-supplied subsets of causal factors. All combinations of models and subsets of causal factors are assigned weights indicating their relevance. Each combination contributes to the final forecast according to its weight.
Therefore, understanding the demand patterns for your products could help you apply the correct forecasting method to the item and location combination in Demantra, which will greatly improve forecast accuracy.
For example, if you already know that there is a product line that shows only intermittent demand patterns, turning off other forecast models for this combination could significantly improve forecast accuracy since the other forecast methods will not contribute to the final forecast number.
Demantra uses the following forecast models:
• Regression
• Regression
• Log (log transformation before regression)
• CMReg (selection of the Markov chain of the subset of causal factors)
• Elog (uses Markov chain after logarithmic transformation)
• Exponential smoothing
• Holt
• Bwint
• Intermittent patterns
• CMReg for Intermittent
• Regression for Intermittent
• Crouton
• Time series models
• ARX and ARIX
• Logistics and AR Logistics
• Other models
• BWint (a mixture of regression and exponential smoothing)
nodal tuning
One reason for poor forecast accuracy for clients using Demantra’s Demand Management (DM) module is that statistical methods and algorithms apply to all combinations or not at all. There is no flexibility to choose specific statistical models for a particular combination different from the rest of the population, although the demand pattern exhibited by that item-location combination may be different from the rest of the combinations. This turns out to be a major limitation during the forecast adjustment exercise for clients of the Demantra DM module.
This restriction is overcome in the Demantra AFDM module which provides advanced analysis capabilities through the Nodal adjustment feature.
Nodal Tuning is a powerful functionality available in Demantra’s Advanced Forecasting and Demand Management (AFDM) module.
Nodal Tuning allows demand planners to choose the statistical models that the engine should apply to a particular combination of item and location to generate the system forecast, and also allows the engine parameters to be set for that combination.
Nodal tuning also allows you to fine tune specific Demantra motor parameters for the combination.
This feature provides a tool in the hands of Demantra experts to fine-tune the engine for greater forecast accuracy. This feature, along with knowledge of the type of demand patterns as mentioned in the previous section, would allow users to enable only those forecast models that fit the demand pattern. This greatly improves the accuracy of the forecast.
Causal Factors/Promotions
Great care must be taken when modeling causal factors in Demantra. If the data model has many causal factors and promotions, they tend to dilute the baseline forecast and result in a highly biased forecast.
A good practice for introducing causal factors into the model is to first start without causal factors and promotions data to generate a baseline forecast from Demantra. Once the baseline forecast is adjusted, other causal factors must be entered one by one, taking into account the effect of any causal factor input on the baseline forecast.
In this way, the effect of causal factors on the baseline forecast can be easily tracked and analyzed, and any time the introduction of a causal factor does not appear to have the desired effect, it should be disabled.
forecast tree
The forecast tree determines which item/location aggregation combination the engine will forecast on. The engine examines each level in the forecast tree and validates whether there is enough historical sales data available for the forecast or whether the generated forecast has sufficient accuracy at that level. In case the validation fails, the engine goes to the next level and continues with the validation phase until it finds a level where it can generate a forecast.
In case the engine ends up forecasting at a higher level of aggregation in the forecast tree, the forecast is split into the lower levels.
The forecast tree is a system setting that has a direct relationship to forecast accuracy.
This is one of the first settings that should be done after careful analysis of the sales history and after discussion with users. Forecast levels should be meaningful to business users and it is recommended to have 3-6 levels that the engine can traverse and forecast.
It is helpful if the forecast tree includes the level at which accuracy is measured, if possible.
relationship
The proportions are very important and are used during the aggregation of the forecast from the lowest level to the highest levels and the disaggregation of the forecast produced at the highest level to the lowest levels.
The final result of the forecast generated by Demantra could be very different depending on the proportions.
The ratios are calculated and stored during the sales history data upload. Several parameters control the calculation of proportions.
One of the parameters that influence ratios is the amount of sales history data that the system uses to calculate ratios. The ratios calculated based on 12-month sales data would be different from those calculated based on 6-month historical data. Therefore, the proper setting of this parameter is crucial for the calculations of the proportions, which, in turn, influence the final forecast.
recommendations
Demantra engine tuning is a complex exercise and there is no single solution for it.
A major engine tune-up exercise should be carried out every two years and whenever there is a change in the demand pattern for the products. The tuning exercise should be tailored to the customer’s specific Demantra implementation, but having an understanding of the factors that influence forecast accuracy would go a long way in further improving forecast accuracy.