The 2 phases of analytics

Analytics means analyzing data to:

1. Discover or understand the behavior of a system or process and describe it.
2. Identify the current condition of the system or process being analyzed.
3. Identify whether the process has changed, or is changing, its behavior with respect to what is known or expected.

Case 1 is known as descriptive analytics, while cases 2 and 3 correspond to what we call dynamic analytics. The following simple, familiar example helps explain the difference between the two.

Example: an inventory problem

Suppose your logistics system consists of dozens of warehouses, each holding thousands of products, and you are responsible for managing them. Among other challenges, you face the following:

1. Managing every inventory correctly.
2. Recognizing that products behave very differently from one another.
3. Defining the management rules that tell you whether a product is well managed.
4. Identifying poorly managed products that cause problems.
5. Knowing whether a product has changed its behavior so its management rules must be redefined.
6. Staying informed, at any moment, of which products pose a risk or problem in the inventory.

Descriptive (static) analytics

The first three points belong to descriptive analytics. What does “managing correctly” mean? One possible answer: no product should be missing when needed, but no product should be held in excess. “Missing” and “in excess” mean something different for every product.

To understand how each product behaves, two associated variables are considered: the quantity demanded over a period of time and the frequency of demand per period. Analyzing them leads to classifying products — for example with the Pareto, or 80-20, criterion — and the classification can differ by warehouse: the same product may classify differently from one warehouse to another. With that classification, an appropriate inventory model is assigned to each product in each warehouse, an association that requires the analyst's good judgment.

Once the model is defined, specific parameters are estimated from historical information, complemented by adjustments based on the analyst's experience: criteria related to customers, suppliers, seasonality, and other information not reflected in the inventory data itself. These are laborious activities, but fortunately they do not have to be done every day: we do not expect significant changes overnight.

Dynamic analytics

Once the theoretically ideal way of managing each product in each warehouse is defined, it is important to watch operations day by day and identify whether something is not behaving as assumed. This answers questions 4, 5, and 6 above and demands much more frequent analysis of the information.

“In a dynamic environment things change: what used to be right may no longer be right today.”

The technical problem is knowing whether a recent set of data — inflows, outflows, inventory levels — is consistent or not with the expected behavior of each product-warehouse. Several mathematical techniques answer with a yes, a no, or a maybe; they can be programmed on a computer and are so fast that, in practice, they can run millions of times a day.

Why analyze new data so frequently? First, because a poorly managed product generates an organizational cost that experience shows to be significant. Second, because in a dynamic environment things change, and it is far less likely that something that was wrong fixes itself by chance.

“The faster we detect something undesired and adjust it, the more we minimize the cost it generates.”

Unlike descriptive analytics, dynamic analytics problems do need to be solved every day, which requires significant computing. Fortunately, these analytical tasks can be programmed and require no human participation to run: they deliver a list of findings of interest — the product-warehouses that meet the task's criteria.

Human and machine

“Humans think and assign meaning; machines calculate and process information.”

Once the findings of interest are identified, people with judgment decide on each case. Human-machine interaction becomes far more efficient: the human thinks and assigns meaning, while the machine calculates and processes huge volumes of information. These analysis tasks are repetitive and computation-heavy but have a predefined interpretation; in some cases it is likely that nothing is found, making them unattractive for humans — ideal candidates for automation.