Anticipate Business Performance with Predictive Analytics

In the mid 1990’s, when the phenomenon of the World Wide Web swept across the globe and ushered in a new era of information, the term “Predictive Analytics” was not yet common industry jargon as it is today. Back in the days when Windows 95, AOL, and Yahoo! were how one could surf the web, would anyone have predicted that we would be able to have this level of digital sophistication to measure and forecast what happens on the internet?

We at Analytics Metal learned early on in our professional careers that there is a significant appetite from the executive suites in organizations of all types that predicting what would happen next could be the game-changer that would transform their industry and make their own organizations more profitable and more successful than their competitors. Predictive Analytics takes the data that you already have collected and tries - sometimes with remarkable accuracy - to precisely forecast what will happen next. In some cases, prediction works wonders and truly helps organizations achieve loftier goals. Other times, organizations fail to budget and allocate bandwidth to analytics departments, or, miss the mark on how to use and apply prediction to help in the business and don’t know exactly what to do. Some organizations can even invest millions of dollars in hiring data scientists, data engineers, and onboard tools like Snowflake, Qlik, Power BI, Tableau, Alteryx…and of course, SharePoint or Jira to manage it all, and never realize one dollar of return on their investment.

The above is assuming that the data that an organization collects is at least in a halfway-decent shape, which is seldom the case and often increases the challenges with Predictive Analytics. Another level of complexity to add to challenges in Predictive Analytics is that data is very often raw, uncleansed, disorganized, and, sometimes, inaccurate.

Prediction Meets Process: The “Analytics Metal” Approach

What software, which models, or what outcomes are decided upon vary greatly by the type of project or task, scope, variables, available data sets, and many other factors in the world of Predictive Analytics. However, the process that we at Analytics Metal employ has proven to work for us over the years, with our clientele and within the organizations that we have worked at in our past lives. This “process for prediction” is as follows:

1. Understanding the Business Objectives, Data Sets, and Key Stakeholders

While it is true that this is applicable to almost any business scenario, it is one of the most important steps in our “process for prediction”. Without a deep understanding of your business objectives, your organizational KPIs, available data sets, and your key business partners and stakeholders, the probability that our Predictive Analytics work will succeed will be low. Analytics Metal will spend the time to learn, to listen, to think critically, and to formulate a plan of attack with you and your team before starting any work (and, certainly before building any models)!

2. Data Assembly, Data Cleansing, and Data Preparation

Much like how a house is built on top of a solid foundation, the team at Analytics Metal spends the majority of the time in a Predictive Analytics project in building a solid foundation with your data by assembling (which could include linking and joining multiple data sets together), cleansing (which could include data formatting, filtering, dealing with null or blank values), and preparation (which could include employing methodologies like outlier detection, classification, and, sometimes, deciding on a temporary or “staging” location for the data). Typically, anywhere from 50% to as much as 80% of a project’s time in Predictive Analytics will need to be spent on data munging.

3. Exploratory Data Analysis (EDA)

An initial analysis, often called an exploratory data analysis (or, “EDA”) will be conducted before model development and selection occurs. This is done so that analysts and data scientists can further develop an understanding of the underlying data set and can begin to formulate hypotheses and strategies for what types of modeling and forecasting will make the most sense for use. We at Analytics Metal like to use the EDA as a critical touch-point between our agency and your organization, presenting initial findings and observations back to you for calibration and project level-setting.

4. Model Development and Model Selection

Many data scientists and engineers consider this stage their favorite, as this is typically the turning point in a Predictive Analytics project, where measurement, scoring and performance come into view. Whether a simple linear regression, ANOVA, ARIMA, decision tree, neural network, or other model is eventually chosen, the process to chose and trial different modeling techniques and the process to evaluate their performance can be very rewarding and fascinating, hence why it is what many in the field of analytics look forward to. However, as the notorious saying goes, “All models are wrong, but some are useful”. We at Analytics Metal will make the best use out of the most useful model and work with your team to explain why and how.

5. Control, Train, and Test Group Selection

Oftentimes, professionals in the analytics space will bounce back and forth between model development, selection, and configuring their data sets for performance testing. In other words, analysts and data scientists will often repeat stages 4, 5, and 6 multiple times in a project in Predictive Analytics, documenting their progress and scoring their models along the way. There is no set number of times that this occurs, but in all cases, we at Analytics Metal will split data into a control group (or a “hold-out” group), a training group, and a test group to achieve high levels of trust and confidence in the predictive power of the models we have chosen.

6. Performance Analysis and Next Steps

Many software programs, like R, Python, Alteryx, and others will provide a number of metrics to help analysts and data scientists score and rank model performance against a training data set and against the control group. These scores are critical to understand, and choosing the right scoring technique is an additional key factor in overall project success. This stage is the home to many acronyms and deep statistical frameworks - while we at Analytics Metal aren’t afraid to get in the weeds and discuss why we use mean square error versus mean absolute percentage error, or discuss the ROC area under the curve, or confidence intervals and prediction power, we tend to keep our results at a high level so that your team can eloquently explain the relevant details of a Predictive Analytics project to others within your organization.

7. Time-Series Forecasting

Once a final model or final few models are chosen, forecasting what will happen after model deployment and presenting those results to your organization is a standard “Analytics Metal” in a Predictive Analytics project. We use this stage as another opportunity to present findings and recommend actionable next steps, keeping your business goals and outcomes in mind and demonstrating net lift and return for your investment. While there is always a degree of error in any forecast, we at Analytics Metal like to have our clients feeling confident and feeling ready to implement the outcomes of a Predictive Analytics project.

Analytics Metal is a practice forged from a blend of robust analytics and deep experience across many industries. Predictive analytics helps organizations go from reporting on historical performance and being reactive to becoming proactive and predicting what is likely to happen next. We would love to help your organization achieve what we have helped others achieve. Let’s chat!