Naturally, to start with predictive analytics, it is mandatory that data is available. Yes you can build nice models and mathematical expressions without data. Yes math allows you to build great algorithms. But NO, you need the data as well to help reaping results. This does not mean, the available data has to be perfect or cleaned. Data Quality companies might try to convince you to clean you data first but that is not necessary. The techniques employed have their foundation in statistics and statistics assumes variations in your data. The only real need is to have data-set available as this helps the communication and exploration of what predictive analytics can contribute.
With predictive analytics, the objective is to create models that predict the future so you can take corrective actions in time. An example can be predictive stock-levels of key components in your manufacturing process. Alternative can be to predict the responses to new product introductions. Closely related to predictions are simulations to help identify the optimal choice to be made.
To employ predictive analytics sensibly, 3 core aspects need to be kept in mind. Together, the help ensure results can be put into production as a clear insight in risks versus returns exists.
LOOK CAREFUL WHAT YOU WANT
Many algorithms exist that can be deployed, whether it is for text analytics of to analyse numerical data. With the availability of so many libraries and open-sourced tool set, it is often forgotten that a specific understanding is needed as well. One can compare it to a DIY store. The large selection available leads to an over-confidence in the projects one can do!
A similar issue occurs when it comes to algorithms. The large set readily available and continuously growing, leads underestimating the need to understand the meaning of the algorithms. It boils downs to understanding strengths and weakness as well as to knowing the context of their validity.
Accurate predictions are only possible with good calibration or fitting. Often data sets are split where one is used for calibration and the other for verification. The calibration process is important as over fitting the models result in ghost predictions that have no bearing on reality.
Also proper understanding is key. An example is the prediction that red will be the new fashion color for shoes. True, the model was well calibrated from a purely numbers perspective. However, from a context perspective, the calibration did not take into account that red was the fashion of last year. A simple example such as this indicates that it is not only important to know how to calibrate models but also what is being calibrate.
Last, it is important to realize the models need to be regularly re-calibrated overtime to ensure they remain valid.
Besides the accuracy as defined by the calibration, it is important to know the quality of the model. A variety of measures exist as indication of the quality of a predictive model.
- A model might only be applicable to a narrow set of situations
- A model might be highly susceptible to changes in specific parameters which might make the output
- None-linear effects due to the coupling of inter dependency of models might result in instabilities in the predictions
These quality considerations need to be considered and clear choices need to be made. Accuracy of the predictions might have to be sacrificed for a more robust model. Alternatively, a large set of models can be used if accuracy is important.
How KENTIVO helps:
KENTIVO can support organisations in leveraging predictive analytics:
- Build Predictive Models: Depending on the need, KENTIVO can build predictive models. This can be done either from scratch or using the library of models we have from work in the past and from our R&D. Naturally, we can also help improve models you made before yourself.
- Create Simulations: Often specific questions are posed, and an organisation wants to evaluate options in a more quantitative way. In this case, KENTIVO will develop a number of simulations and analyse the impact of different options available. We also will help explore new options and ideas in this way.
- Help Apply Tooling: Regularly, we encounter organisations that have already procured tooling. Often the reality in implementing afterwards less straight forward than anticipated. Or, the tooling needs to be employed in new domains. In those situations, we support an organisation in getting more out of there toolset/workbench.