Why do analysis? This may seem self-explanatory, we want to answer questions, that's the reason most analysts do analysis, but exactly what questions are we trying to answer, and what type of analysis do we do to answer those questions?
There are four categories of analysis, as defined in Gartner's Analytic Ascendency Model:
- Descriptive analysis, which describes what happened in the past
- Diagnostic analysis is what we do when we want to find out why something happened
- Predictive analytics gives us insight into what is likely to happen, given past behaviour and conditions.
- The final step in the value chain is prescriptive analytics, which tells us what we must do to achieve a specific outcome.
Gartner's Analytical Ascendency Model
Descriptive Analysis
Assuming we have the data (and that is a whole topic for another day), the questions that are typically asked start with "What happened?" Analysis that answers this question is called Descriptive analytics. Descriptive analytics makes use of historical data, and it tells you what has happened in the past. Descriptive analytics are used in reports and are typically summarised so that the reader can consume and make sense of the information.
Descriptive analysis is typically used for reporting, to understand how things may have changed over time. This type of analysis is also used in benchmarking exercises and understanding seasonal patterns. In many ways, this is the simplest form of analysis, it is reasonably straightforward, people understand that it shows a picture of what happened in the past and it delivers information to decision makers.
Diagnostic Analysis
If something out of the ordinary happened historically, typically the very next question is, "Why did it happen?". This brings us to Diagnostic analytics. Diagnostic analytics is all about understanding why something happened, and is typically a more detailed process that requires more data and context than required for descriptive analytics. In retail, holidays that shift annually are a red flag, for example Easter. In some years Easter and the associated long weekend happen to fall in March, whilst in other years, Easter falls within April. If you're wondering why April sales this year are higher or lower than they were in previous years, having a good understanding of when the Easter weekend happened could be key to answering this question. In addition to the shifting holidays, it would be good to understand what promotions or incentives were being offered. Understanding context, in addition to having the data to unpack the context, is key to good and robust diagnostic analysis.
Predictive Analysis
Once we know what happened and why it happened, we often ask, "Knowing what we know, what will happen in the future?", and this brings us to predictive analysis. Predictive analysis is exciting, we're looking forward into the unknown and asking what will (or is most likely to) happen tomorrow, next week or next year. Essentially, analysts use modelling techniques to analyse historical data, and develop models which can then be used to predict the future.
Techniques typically used include logistic regression, regression or classification trees and random forests; to name a handful of the more common ones. It is important that analysts understand the strengths, weaknesses and limitations associated with these techniques so that they select the best technique for the problem. In addition to the technique used, there are considerations regarding training and validating models, feature engineering and target variables. Producing robust predictive analysis is a highly sought after skill.
Prescriptive Analysis
This is an advanced form of analytics, and is used to determine how to achieve an outcome given a set of constraints. Techniques used in prescriptive analysis include Markov Chains, mathematical modelling and Monte Carlo Simulations.
Markov chains effectively allow us to step through time; a Markov chain gives the likelihood or probability of a specific outcome or event happening at a predetermined point in the future. A Monte Carlo simulation is a technique where input variables are modified based on probability distributions such as the normal or log normal distribution. Simulations, or iterations, are run, using these modified inputs; the likelihood of the output is then calculated, based on the modified inputs and the model. This is technique is most commonly used when a model's inputs are uncertain or variable and the model is complex and dynamic. Mathematical modelling uses (unsurprisingly) maths to define and describe a system that consists of input or feature variables which are typically subject to constraints, for example the maximum capacity associated with a call-centre; as well as an objective, the desired outcome, which could be maximising $ collected, minimising losses, or could be a combination of objectives, such as maximising approval rate whilst minimising bad debt.
Hopefully this has given you a flavour of the different types of analysis that analysts can do, from the most basic, understanding what happened, to the significantly more complex forms of analysis. The fundamentals of technical analytical skills are straight-forward, but as we become more experienced, we can continue to improve our technical skills, ensuring that we add additional value to our teams and organisations.
Where are you at? Are your analytics primarily concerned with describing what happened, or do you spend time playing in other spaces? I would love to hear about your experiences, what works for you and what challenges you may be facing.
Further Reading:
Comments