Visualizing Schedule “Tradesies”

We recently met with an operations executive that was confronting one of the more vexing issues a manager faces. An employee told her there is a rumor going around that employees are swapping shifts with each other to increase their pay. In this organization, anyone who is not scheduled to work but then works is entitled to premium pay.

One of the well known ways to exploit this type of policy (I explained this as “tradesies” in my book Lean Labor) is to find a couple of buddies and regularly swap shifts with each other. It may not be every shift, but you have to trust your partner enough to know that you can be up or down a shift and when the time comes, they will agree to swap out a shift with you.

This of course drives up costs without increasing output. Not a good outcome for any organization. The executive was pondering what to do. The challenge is that there are many legitimate reasons to swap shifts and the policy is intended to provide flexibility for the workforce but ensure coverage for the workload. A premium may be paid to encourage employees to work hours that they might not otherwise want, thereby providing liquidity to the system.

Addressing this would be tough because it’s difficult to discriminate between legitimate shift swaps and ones that were done purely to increase pay. But the rumor was expanding and if this practice spread it could ultimately lead to lower profits, poor morale and even layoffs of uninvolved people.

Before she acted, she asked my team to take a look at her organization’s scheduling data and see what we could find. This is a fairly challenging exercise because first you have to figure out who actually swapped shifts with who. There is no “marker” other than a premium paycode for one person. After that was resolved, we had a long list of shift swaps. Next we had to figure out a way to visualize that list to help interpret the data.

After a couple of different approaches, the team was excited as they realized this would call for a different type of visual approach. The reason they were excited is that the vast majority of visualizations required are bar, line and scatter charts. These charts do a great job, but we all like some variety!

In this case the team realized they were looking at a networking relationship between the people swapping shifts.

Using a networking diagram, they plotted the employees and who they swapped a shift with. What we wanted to know though was not only who, but how many times shifts were swapped since gaming typically occurs between a small group of people. For that we colored the arrow differently based on the number of swaps made over the time period analyzed.

schedule swap diagram

Below is the result of the effort. As you can see, the majority of the swaps are occasional and with a variety of people. Good news! Most people are swapping shifts as the organization intended. But after applying a filter to remove the occasional swaps there are two clusters of three people that are swapping significantly more times and with the same people. This doesn’t necessarily mean they are gaming the system. It’s possible that they have very specific skills and there is a limited pool of people they can swap with.

This information was illuminating for the executive. Out of thousands of people, she could now focus on six and get to the bottom of it quickly. She could also respond to the rumor with hard facts. Finally, it was peace of mind for her to know that the vast majority of her employees were using the policy as it was intended.

The data scientists behind the scenes and how they put a spotlight on dark data

Over the past 18 months I have learned a lot about analytics and big data, especially applied to the workforce. I spend a fair amount of my time speaking to customers, analysts and the media and besides the most common question of “do you have any examples you can share” I get asked about the people involved and process of developing these applications. In the spirit of “people are the most important resource a company has” I want to showcase that side of big data.

For timekeeping and scheduling, dark data…( to save you a quick trip to your favorite search engine, I mean data that is collected but not typically used, yet is still required) ….in this case the audit trails of any change made to a timecard or schedule.

In Kronos there are sixteen different types of edits you can make to a timecard or schedule. Each of these edits represents tiny trades of time and money between an employee and the company. Individually most are inconsequential, but in aggregate they represent tens or hundreds of millions of dollars. By and large most of these changes are transactions that everyone agrees to and are necessary… The employee forgot to clock in so the supervisor adds in an “in punch”. Or an employee calls in sick and the supervisor changes a paycode from regular to sick in the schedule.

Occasionally however there are situations where the changes are indicative of an issue. For example, a supervisor changes a couple of minutes around during the week on an employee’s time card and eliminates premium pay. Or a supervisor changes a schedule after the fact to represent that an employee only worked the hours they were scheduled.

These small changes are usually lost in the millions of annual transactions that occur throughout the year. And because they are so small they are usually missed by most reports and audit teams. Only when the employee affected has the courage to speak up does a company become aware of it. By this time the consequences for all involved are significant; from degraded morale on the part of the employee to unnecessary cost in terms of productivity, turnover and financial impact for the company.

As the economy improves and companies feel the pain of turnover and lost performance when employee engagement sags, we have been engaged by companies to understand how they can identify these situations. The companies know the answer is in the data because when someone files a grievance and points out the specific situation and dates, the HR department can immediately see what happened in the transactions.

The challenge is seeing these changes sooner; especially before someone is so frustrated they file a grievance or the behavior becomes obvious to all. This is where one of our data scientists who has a PhD in computer science realized that this is a very similar challenge to what retailers face when they are trying to understand what the millions of customer clicks represent on a website. The customers aren’t telling them why they are clicking the way they do and only a fraction of the clicks result in an order.

So the data scientist applied the same machine learning techniques on timekeeping data that retailers use when they analyze their web server logs. The result of his work however was very difficult to interpret unless you understood machine learning and clustering techniques. To simplify this we had one of our visualization experts re-imagine the output in a way that a lay person could understand. Her interpretation was amazing in its simplicity!

Secondly, the data scientist had created a very flexible tool. The first prototype had a number of tuning parameters requiring the user to take output from past results and enter it in to help weight certain parameters for future analysis. We recognized that aside from a data scientist, we couldn’t expect a typical business user to be able to perform this tuning. So we focused what the tool could do and eliminated the tuning parameters.

clustering dashboard

An example of the machine learning dashboard in Workforce Auditor

We were very nervous and excited about analyzing our first data set (we went in without knowing anything about the customer or their practices to ensure we didn’t bias the analysis). When we researched the results, there it was…we had found an issue that was previously unknown to that company. We tried it a second, third and fourth time. Each time we found something important to the customer that they suspected but couldn’t prove or they were completely unaware of. These small changes were indicative of million dollar +  issues that were looming for these companies but had now been avoided….very exciting stuff.

We found supervisors gaming schedules to improve their own bonuses (the company since tweaked the rules of the bonus). We found a store manager working extremely hard to rebuild her schedule each week because the forecast and automated schedule she received was off (the company immediately re-tuned the forecast for her store). There were many more examples and we realized that we had developed quite a versatile tool. Its power is that it can evaluate the actions of thousands of employees and narrow it down to just a handful of situations that require further investigation in a matter of minutes.

With so many positive results we fast tracked the technology. It’s now available as Workforce Auditor and is included with our Workforce Analytics platform.

Take a ways from this experience?

1) Skills and experience really count in developing big data applications, no one is going from “excel guru” to building a machine learning application overnight and it takes multiple people to get it right

2) involving (internal or external) customers and their data is essential; no one could ever build this without deep domain knowledge and many different data sets to trial

3) By focusing on the business problem rather than the technology we created something that was streamlined and easy to use rather than a feature laden product showcasing the power of machine learning.

When I have a little more time to write, I want to share how the newest member of our team used scheduling data and a network map to uncover undisclosed relationships in a company and what it was costing them….stay tuned!