Running, Heisenberg and flight risk.

cross posted on my work blog)

Readers of this blog will know that I am an avid but plodding cyclist.  It gets me away from the desk, and means I can bore for Germany on components, frames, cadence, altitude and the like.

I sometimes run, and recently an old school friend bet me that I couldn’t run 10K under 50 minutes by the end of September. Not being one to ignore a challenge, I took him on.  

Running used to be a simple affair.  But in order to go for a run, I “need”

a) my minimalist five fingers running shoes.

b) a  fully charged iphone

c) headphones that don’t fall out my ears

d) heart rate strap

e) strava run application.

f) the right spotify play list.

Measurement has changed running. I know exactly what heart rate to run at and at what pace per minute.  It takes me at least 10 minutes to get out the door.

Heisenberg’s uncertainty principle is one of the most abused scientific principles, but it applies here.  My efforts to measure running has changed my running.

What has this got to do with HR systems?

I have noticed a  trend from vendors to include a field called Flight Risk into their talent management systems.  I reckon this isn’t a good idea.

1. 99% of HR systems don’t have accurate enough data to remove false positives

2. The data to really predict flight risk isn’t in the HR system.

3. Telling someone who isn’t a a flight risk that the system thinks they are one will make them into one.

4. Managers will react differently to the same data.

5. It will taint other, more accurate results.


To assume you can build an algorithm to predict whether a person is likely to leave their job or not based on the shoddy data in a talent management application is arrogant and irresponsible.  Vendors’ crude attempts to measure human intentions will create unintended consequences, most of them bad.


In case you are wondering, I won the bet…




3 thoughts on “Running, Heisenberg and flight risk.”

  1. Thomas,

    Great post. Flight risk is very closely linked to customer churn. What you’re trying to do is improve your accuracy of prediction using available date above the level you’d naturally give. The comparison group is likely to be the ‘informed expert’ namely the manager.

    All predictions have a probability. To make a decision to act you need to incorporate a loss function (where software tools usually fall down, partly because it requires a user choice). The loss is asymmetric and potentially great for employees, much like drug testing cyclists. Sports bodies choose to minimise false positives, the counter side of this being they produce a large number of false negatives.

    Can you build prediction with just the information from one system – of course you can, you might just not get much better results than asking your expert. What any such system should have is a mechanism to review accuracy over time.

    In marketing, when looking at churn you use the loss function to determine how expensive your offer is. You want to make sure that the benefit of reducing the churn is more than the cost of the offers you have to give. For employees you want to do the same. In churn you’re probably working with single-percentage-figure improvements

    My view: this analysis is potentially useful. It’s probably most useful if it reports the factors that influence flight risk as factors can guide action. You want to ensure that it’s not a black box because some of the best predictors might be factors that you can’t act on (like gender, race, age). I like decision trees for this type of work as you can explore the reasons.

    If you do want to do prediction then it’s useful to roll-up binary classifications to a department or skill-group level to have an indication of the likely turnover in the period you’re interested in. This can be fed into a planning process.

    Bottom line: get an expert to do it and guide you through how to use the results rather than use a system.

    For an example of how we’re building prediction into a large number of dashboard types for clients see here:

  2. Hi Thomas, I would consider using a flag to indicate employees critical to my organisation rather than marking those who are a flight risk. For one, it sends positive message (agree with your view that knowing that he/she is a flight risk will turn an employee into one. Similarly, knowing that one is critical for an organisation, should make the employee more engaged).
    Second, we can focus on doing the analysis to check if the cost of churn is greater than the cost of offer can be limited to these employees.
    Third, my talent management and succession planning definitely becomes easier.
    Peoplesoft has this field. I haven’t noticed this in other products.

  3. So what other data, not in the HR system, would make sense to look at?

    Could the pain of incorporating that non-HR data be justified?

Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out /  Change )

Google photo

You are commenting using your Google account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s