Tailoring analytics to your needs

Automated Decision Support Systems

Senior decision makers, independently of the type of organisation, have the same set of resources available:

  1. Data information systems
  2. Industry experts offering their insight
  3. An analytics platform often referred to as a support system

The ability of the organisation to make the right decisions relies, amongst other things, on a good interaction between the above resources. For several decades Decision Support Systems (DSS) have assisted organisations in improving their decision making processes. The effectiveness of a DSS relies on the skills of the users, especially their experience with using the tool. However, the dependency on the users' expertise can be avoided with Automated Decision Support Systems (ADSS), where decisions are automatically taken by a rule-based system. Through advanced analytical techniques we can avoid the rules becoming outdated and we can ensure the model is continuously learning from new data without requiring a user to manually update the model. We often refer to these techniques as mathematical models using Machine Learning (ML) or Artifical Intelligence (AI) algorithms. When incorporating self-learning models, the organisation should introduce intelligent controls to monitor the behaviour of ADSS in real time, which can be achieved through model backtesting techniques designed for ML and AI models.


The following developments have been instrumental for advanced analytics to become accessible and widely implemented.

  1. Significant improvements in Machine Learning and Artificial Intelligence open source libraries.
  2. Parallel calculations on multi-core CPUs and GPUs have reduced the calculation times to allow for continuous improvement of analytical models.
  3. Access to large volumes of memory and increased speed of memories has enabled organisations to collect and utilise more data.

As an organisation that embraces advanced analytics, we utilise these technological advancements in our analytics platform. Our analytics platform is intended for modern businesses to quickly understand the value of their internal data and improve their insight. Our application design choices, such as using open source software, permits us to integrate the platform on various operating systems and hardware to scale the platform to the need of a client. The goal of the platform is to automate decisions in order to improve their consistency, increase their transparency and improve the speed with which they are taken. Given that decisions are automated, we ensure the models automating decisions can be monitored at the same frequency with which they are updated. For instance, a model is rebuilt at the frequency of newly available data and simultaneously backtested.

Binary Decision Model

Often managerial decisions can be broken down into binary decisions. For instance, complex problems can be translated to "yes" or "no" questions such as

  • "Do we think we need to invest in this project?"; or
  • "Does the organisation need to increase FTE?".

Decisions need to be validated and require transparency for stakeholders to understand the reasoning behind them and to be able to follow up on the impact they have on the organisation. Often a decision is evaluated via meaningful metrics prior and post decision, such as KPIs that drive the productivity of an organisation. In advanced analytics we look for raw data that can be engineered as features in such a way that they can easily be understood by mathematical models. In other words, we transform raw data so correlations between metrics and past decisions become evident.

Our platform can quickly evaluate the information of a given data model and verify whether it is explanatory of past decisions. After a data model is build, we can output the model to use it in other environments or for audit trail purposes. Furthermore, we can easily run scenarios and predict the impact of modifying certain features, e.g. reducing the FTEs of a certain profile while increasing FTEs for another cohort of profiles.

An example of binary modelling is predicting the probability of a credit being either performing or default, which translates to evaluating the creditworthiness of a client or a group of clients based on their creditworthiness features. Such a model can be useful when evaluating whether you should underwrite a credit.

Note that in this example not only the prediction of being defaulted or performing in the future is important but also the likelihood of being in default. Often, not only the most likely event but also the likelihood of a given event is meaningful and will allow the organisation to investigate which scenarios of features drive an outcome more or less. In other words, ranking probabilities is an important part of ADSS. The ranking power of probabilities predicted by a given model is an example of how to evaluate the performance of a model.

In order to easily rank probability, it is conventional to use classes or ratings. The platform can automatically calculate ratings based on the granularity, i.e. number of rating grades, the organisation would like to use. The impact of translating probabilities into rating grades is evaluated in the backtesting of the binary decision model.

The backtesting focuses on the model's ability to differentiate between classes, i.e. different outcomes/events, and the accuracy of the predicted probability and the actual probability.

Multi Response Decision Model

Not all managerial decisions can be expressed through binary choices, and in such situations we require multiple response classes. For instance,

  • "Is the productivity of a plant low, medium or high?"; or
  • "Will the fund manager generate a negative alpha, no alpha or a positive alpha?".

Aforementioned, ADSS require features, i.e. data inputs, that are engineered to explain response values, i.e. classes, to a particular question.

Decisions are often based on the expected outcome of a given event. A good example is Loss Given Default (LGD) models that aim to predict the loss on a defaulted credit. The assessment of expected loss for different scenarios determines the decision of the organisation's recovery strategy: for example, whether the sale of collateral should be done internally, externally or as a joined effort between internal and external resources. Depending which scenario projects the lowest loss, a decision can be taken automatically. Note that on our platform we model finite classes, i.e. events, not continuous values, therefore, in case of LGD, we will model its distribution based on discrete loss intervals and not continuous loss values.

The different classes being modelled can, in certain cases, have a meaningful estimate, which is the case for LGD. For every class we can calculate the expected values and, when we model classes, we can either look at the predicted class or the weighted probability of each class' estimate. In our backtesting tool for multi response decision models, we expect both the class and the estimate to be meaningful. Therefore, we evaluate both the accuracy of the predicted estimate and the linear dependency between predicted and observed classes.

Calculation Platform

After building our ADSS, we want to be able to use it and, equally important, understand it.

The platform provides a functionality to built and save models that use Machine Learning algorithms or Artificial Neural Networks (ANN). After saving the models, we can upload them to the platform and apply them to new samples. The model output for the new sample can be downloaded as a flat file which is easily read by excel or similar programs. Users can investigate what features were used by ANN to evaluate a response value within the sample, which improves the transparency and the understanding of ANN.

In the case of LGD models the platform can be used to understand which features predict the likelihood of a certain loss scenario.