The assessment and monitoring of changes in the actual values of key indicators allows us to determine the level of achievement of the specified strategic goals, but it is frequently necessary to investigate the dynamics of their changes in future periods in order to build an analytical foundation for making management decisions.
Sapiens was a project partner in terms of developing a forecasting methodology and automating the process, which began with data gathering and ended with outcomes display.
The characteristics of the projected indicators are such that market elements such as commodity and currency futures dynamics, parameters describing regional meteorological conditions, and macroeconomic indicators impact the change in their values. The values of the anticipated indicators fluctuate as the dynamics of the contributing elements change, which is especially true for long-term forecasting.
External market variables impact projections to various degrees, resulting in a value that is not constant over time, as well as the values of projected indicators.
A multifactorial analysis is carried out with the allocation of the key components and the calculation of impact coefficients in order to discover and incorporate the most significant aspects in the model in the present time period. Algorithms are built for the automatic analysis and allocation of the corresponding components of cyclicity, seasonality, and trends in the behavior of time series from previous periods for statistically justifiable use on the forecasting horizon. The forecasting horizon is a model parameter that may be modified by the user to meet the demands of the organization.
When establishing the methodology and building the solution, it was noted that the actual values for the indicators are received daily, that the identified primary parameters of the forecasting models are determined at each launch by training models on new data, and that the models react reasonably fast to abrupt changes in the dynamics of the projected indicators.
To reduce forecasting inaccuracies, the approach of constructing ensembles of models is utilized, which is intrinsically similar to "voting– a weighted average estimate is derived taking into account previous data and the history of predictions of each of the models, resulting in the requisite accuracy. Algorithms for the formulation of consensus forecasts are established in order to take into consideration expert opinion, which may be provided by responsible personnel from relevant departments.
The creation and execution of the solution (its plan is depicted in the diagram below) occurred in stages, with the following works completed:
- Data was gathered, and an examination of the impact of external market forces on the forecasted indicators was conducted.
- Methods for forecasting have been developed;
- The feature of transferring source data and forecasting parameters from SAP BW to AP Deductor (SOAP) and communicating forecast values of indicators from AP Deductor to SAP BW was implemented, taking into consideration the idea of a layered scalable architecture (LSA).
- Mathematical models with machine learning aspects are implemented using the Deductor Ent analytical platform, according to the methodology. The models may be configured in a variety of ways, both manually and automatically, with parameters selected based on specific criteria; AP Deductor Ent is a reasonably functional tool for the software import substitute program.
- The transaction workplace of the predictive model administrator in SAP BW has been developed, where flexible model management is carried out and it is possible to start work and select parameters "at the touch of a button".
- Reports on predicting accuracy and indicators were shown on information panels (using SAP BEx Analyzer).
- Operational forecasting with the essential precision for business, taking into account changes in the dynamics of indicators on a daily rising volume of historical data, allowing for faster adoption of a controlling impact on the enterprise's activity;
- The findings are interpretable due to the openness of the process at all levels of forecasting - gaining knowledge of the influencing elements and the capacity to back up the conclusions reached;
- Analysts and methodologists in the operational organization will be able to focus their efforts on improving the methodological basis by minimizing the time spent on regular data preparation and recalculation of models using improvised methods, hence lowering the risks associated with the human component;
- Simplicity and ease of use - if the forecasting process must be started manually, the launch is conducted on the screen of the forecast model administrator's workspace, with reporting on each step of the models, according to the type "all computation at the touch of a button."