The stages of a forecasting software implementation project
Artificial intelligence and forecasting projects follow a pattern that can be described through an agile and iterative process. Therefore, forecasting application creation projects cannot start without the definition of a use case fulfilling a precise business objective, such as sales forecasting or stock management optimisation. Project management requires choosing the data to be used to create and train test forecasting models. Once validated, this stage leads to the deployment of the prediction models on the perimeter defined upstream of the project or in the specifications.
The implementation of forecasting software or applications also requires the creation of an exchange base between the client company's business and IT teams and the team of data scientists in charge of forecasting modelling. The planning of this type of project therefore takes into account the business and technical infrastructure to which the final product will be attached. In addition, other project-specific parameters must be considered before the start or during the implementation of the proof of concept, namely the identification of risk factors and potential hidden costs.
Here you can learn more about the cycles and processes involved in setting up a forecasting project, detailing the various constraints that need to be taken into account to ensure that your company's business objectives are met.
Definition of the business issue
Use cases and business objectives
SaaS forecasting application projects operate along the same lines as the rest of the data science projects. The start of these projects requires the specification of a business challenge or business case. This business use case is then converted into a mathematical model by teams of data scientists. The business cases covered by the solutions developed by Verteego can be grouped as follows:
- forecasting of permanent or promotional sales, cash flows, and attendance in physical or virtual points of sale;
- optimization of product assortments in physical stores or e-commerce;
- allocation of human resources, logistics, raw materials or cash;
- Support for new product launches or new collections;
- optimization of storage costs, pricing;etc.
A project may have different business objectives. Verteego ensures a progressive deployment of predictive systems that meet the needs of each use case by prioritizing projects upstream.
ROI-oriented approach and choice of KPIs
Once the use case has been defined, the next step is to evaluate the historical data of the company as well as the applications or tools used in the past by the company to respond to the same business issue, if any. The goal of this approach is to define the key indicators or KPIs necessary to validate the pilot project and to define the variables and data that can be used to model the forecasts. The tools previously used by the company can be used as a reference to quantify the performance of the Verteego tools and their effectiveness in reaching your business objectives.
The values tracked to assess the performance of the algorithms and predictive methods can be the increase in the availability of items and products online or in physical stores, the growth in the number of sales, the positive evolution of the conversion rates, etc. Decreasing metrics can be targeted in the case of objectives to decrease storage costs, delivery times, stock-outs or losses of perishable products. Finally, the return on investment (ROI) allows the success of the predictive process to be quantified and evaluated from the pilot phase.
Pilot project launch and data management
Definition of the test perimeter
In artificial intelligence projects, the POC serves as a pilot project. It concerns a defined perimeter between the supplier of the application solution and the client company. The implementation framework of the pilot use case can, for instance, concern a certain number of points of sale (physical or not), certain products, product ranges or collections of articles. The test project allows the implementation of a first version of the predictive model. The success indicators of the generated predictions validate its relevance.
The pilot use case creation phase lasts three or four months and allows us to quickly draw conclusions on the reliability and quality of the data used and the validity of the project. It also allows Verteego's teams to begin interpreting the prediction lines obtained based on interactions with the client company's business teams.
Choice of data and resources to exploit
The design of the first forecasting model is done in several steps. The first is the selection and pre-processing of the data extracted from the company's data lake (cleaning, elimination of duplicates, standardization, etc.). In addition, during this phase, the business teams and data scientists can verify the availability of data essential to the success of the project and think about solutions to compensate for any lack of information useful for modeling. Indeed, known data science issues may emerge at this time, such as the discovery of a lack of continuous historical data or low or non-exploitable data quantities, etc.
Additional resources are selected during this phase to complete the company's historical data. Indeed, the training of predictive models requires the crossing of internal and external data in order to better explain the variables under study and to optimize them. The sources of exogenous data can be databases containing meteorological or econometric information, information about the competition or characteristics about the geographical areas of the physical points of sale for example.
Production launch and connection to the company's IS
Deployment and connection to the company's IS
Once the pilot project has been validated, the production or industrialization of the resulting model can begin. This involves connecting the platform carrying the forecasting software to the client company's information system (IS) by using the Verteego API, among other things. As the POC took place on a limited perimeter of the company's technical infrastructure, the production launch allows for its expansion as well as the scalability of the predictive model created. The production launch also takes into account the interoperability of the forecasting application with your organization's software packages and applications dedicated to the supply chain.
Improvement of the predictive modelOnce deployed, there is an opportunity to improve the performance of the original forecasting model. This phase can include adding new data or resampling it. The tuning of the hyperparameters of the used algorithms will also serve to optimize the predictive performance of the forecasting software.
This third step also serves as a framework for defining the conditions of use of the platform and for thinking about the interaction between the SaaS platform and the users of the operating company. This involves creating dashboards adapted to the needs of your organization's business or IT teams.
Monitoring and optimization of forecasting over timeOnce the deployment of the application has taken place on the perimeter concerned by your business issue (physical, web or mobile points of sale for example), you can continue to benefit from Verteego software improvements as well as the monitoring of the results of predictions generated by the SaaS forecasting product. Continued support and monitoring is provided on an annual subscription basis.
Verteego and the governance of your corporate data
One of the aspects taken into account by Verteego's teams during the project management (testing, deployment and follow-up phases) is to answer questions related to the transparency of the AI designed to meet business needs. Verteego ensures that the role of the implemented algorithms and the different techniques used (machine learning and deep learning approaches) is explained throughout the application deployment.
Moreover, artificial intelligence projects and, more broadly, business intelligence, projects require attention to European regulations and those of your company's host country (GDPR, HIPAA, for instance). The data capital of companies is therefore subject to both legislative and regulatory constraints, but also to a data management policy defined within the company itself. Verteego defines upstream and in collaboration with your structure the storage and exploitation modalities of your data in order to ensure the respect of your data governance policy.
In conclusion, projects based on machine learning and deep learning methods follow an iterative and progressive process oriented by a business objective. The definition of the project schedule takes into account the cleaning and pre-processing of the data, a test phase to validate the viability of the model to start the deployment of the predictive solution and its connection to the technical infrastructure of your company. Different metrics determined according to the initial business case allow to quantify the efficiency and the quality of the predictions generated by the predictive model throughout the project management. Finally, one of the key steps in the management of an artificial intelligence project is the analysis of how the company's data will be used in order to respect the framework set by its data governance policy.