We’re starting a new feature called Tech Tips, designed to quickly get you that little bit of HR Analytics tip you can’t seem to find anywhere else. Today, we talk about selecting performance data for your predictive model.

Question

A reader recently asked ” If I would like to use the previous year’s data to predict individual performance, should I use only performance from last year, or should use data from previous years too? For example, I would like to predict employee turnover in 2018, should I only use 2017, or should I use data from 2016, 2015, etc?”

Answer

When thinking about previous performance to predict something about next year, I generally stick with only the previous years’ values.

There are a few reasons.

First, the most recent data is likely the most informative; I would rather know about your most recent performance than your performance two years ago.

Second, you will not have more than one year of performance measures for those people hired within the last year. Using just the most recent year puts everyone on the same footing.

Third, just using the single, most recent year of performance data makes my modeling work easier. Knowing about your performance from a few years ago might be helpful but it might not be worth the extra effort (at least if you are just gettting started with HR Analytics).

If you still want to include performance from earlier years, then you will need to think about how to fill in those missing values for new hires. One option is to just use the average performance value across the whole population to fill in the new hires’ missing value. I’m not a big fan of this move but there is no law that says you can’t do it.

Note: My answer here assumes a traditional, once-a-year HR performance rating. More complicated and more frequent performance scoring approaches might require a different set of considerations.

Tips for Building the Predictive Model

Accordingly, I would develop and train my model by using the 2016 performance data (and other variables) to predict 2017 performance.

Given that model, I would then plug in my 2017 data (that is, the most recent data available) to predict results for the upcoming 2018 performance period.