Predictive Analysis is a process of collecting, transforming, cleaning, and modeling data with the goal of gaining consensus of a future forecast. This provides the required insights for making informed decisions. This process consists of the following phases that are iterative in nature.

1. Define The Need Of The Forecast

The very first step consists of understanding the who and why. This could be as simple as a recurring need as part of a monthly demand planning cycle where you generate an item level forecast for operational planning, or it could be an ad hoc request for analysis of a potential new product launch. Whenever we are required to carry out analysis, we first need to know the internal customer, assess the requirement, and determine the required data. Then you can properly do your forecast and analysis.

Then we assess the required resources and return on detail of analysis. Depending on the situation, this could lead to a quick naïve or judgmental forecast or a detailed probabilistic analysis for another. This phase could also could determine what items or customers need to be reviewed. Doing customer-product segmentation prior to the demand planning implementation is a great help in this respect.

Key questions to ask include:

  • What is your time frame? (e.g., monthly, weekly)
  • What level of aggregation is required? (e.g., customer, item, location)
  • What is your unit of measure? (e.g., dollars, units, hours)
  • What factors should be included? (e.g., market data, sales input, time series sales history, causal inputs)

2. Demand Analysis

Data gathering (update historic demand, new data sets)

Data cleansing

The next step consists of gathering and preparing your data. We need to select data as per the need, clean it, construct it to get useful information, and then integrate it all. As part of a monthly demand cycle, this is often just updating historic demand and possibly pruning outliers or for promotions.  It could also mean having to collect entirely new data sets for new analysis or existing analysis. Whether it be an existing data set you are only updating or entirely new data, a critical step is analyzing and understanding your data, ensuring it is in the correct format, and cleaning your data set.

As you collect and organize your data, remember to keep these important points in mind:

  • Whenever possible and feasible, visualize your data.
  • Record data in the same terms as needed for the forecast.
  • Before you collect new data, determine what information could be collected from existing databases or sources on hand.
  • Keep your collected data organized in a log with collection dates and add any source notes as you go (including any data normalization performed).


3. Predictive Modeling


Model selection and estimation

Forecast generation and baseline forecast

Review outputs and add assumptions (gather business intelligence market & customer)

Develop consensus


 Once data is gathered, we can begin the predictive analysis on the data. In the most basic terms this could be generating a baseline forecast, reviewing it and adding assumptions, and then developing a consensus demand plan. For this, we need to select a modeling technique, generate test design, build model and assess the mode, and gain alignment. Whilst most see this starting with a statistical forecast, best practice is to begin modeling with an estimate based on the model and requirements from the very first step of defining the need. Then when the various methods, techniques, algorithms and analysis are placed together into a model and run, we can review them. For this you may consider hold out periods and ex post forecasting and reviewing with key stakeholders.

We are working towards alignment and a one number forecast or predictive analysis. It is common from a baseline forecast to add additional causal or judgmental inputs based on business intelligence, market information, and internal or external collaboration. It is important to rely on a quantitative baseline forecast first – statistics should always be your starting point rather than the finishing post. That said, most models will not have crucial information about product launches, substitutions, and end-of-life products, or other types of relevant information including promotions, price changes and marketing campaigns.


4. Manage Output


Communicate (interpret results)

Manage demand (S&OP demand reviews, demand management)

Measure and track


You can create the best predictive analysis and forecast but it won’t matter if it is not properly communicated, utilized and measured. Our process does not stop at the water’s edge – we need to ensure we met the initial need of why we generated the prediction and continually monitor and improve. It starts with being able to communicate forecasts in the language and output that your audience needs.  As I have said before,  good demand planner are story tellers who use numbers as their language.

Good communication is timely, useful, consistent and formalized. For a monthly process, I would recommend as the final step a Demand Review that may be stand alone or part of a formal S&OP, FP&A or Business Efficiency Process (BEP). It is in these meetings where you can develop collaboration for inputs, communicate efficiently the outputs and any uncertainty in the numbers, and measure the success or shortcomings.

To this final point, any good process should be measured and tracked not only to see if it meets the objectives but to improve future predictions. Predictive analysis and forecasting should be viewed as a continuous improvement process. For this, metrics like MAPE or bias are not just the goal post of performance but should be used by practitioners for insights. In addition, consider using metrics such as Forecast Value Added to help measure the process and inputs for even greater insights. Tracking and measuring forecast performance and the process is an essential part of the forecasting process. If you cannot assess your current process, it is very difficult to improve it.