Offer forecast applies machine learning to the recruiting process to make a forecast for each job about when an offer might be accepted and when a new hire's start date might begin. This model analyzes huge amounts of data and finds complicated relationships to generate predictions that can change over time.
How offer forecast works
Offer forecast is powered by a machine learning model called LightGBM or Light Gradient Boosting Machine. This model is trained on input variables from filled jobs. Some of the data used to train the model include the following:
- Number of candidates in each milestone
- Steps in the hiring plan
- Months an organization has been using Greenhouse Recruiting
- Days a job has been open
- Days since a job was last updated
- Offer status, if any
- Average hiring speed
- Time to fill data from previous jobs
- Lots of other things!
By observing historical metadata, including millions of applications across hundreds of thousands of filled jobs, offer forecast derives a statistical model of job pipelines and candidate movement. It then uses this model to predict future outcomes, like when someone might accept an offer.
Next, offer forecast derives an output variable called time to hire. After many iterations, the model adapts to what input variables are more or less helpful in various situations for time to hire predictions. With sufficient training of the model, it then uses new input variables from live jobs to make an offer forecast about your open job.
How specific forecasts are generated
While it’s not possible to tell you exactly how an offer forecast generated or how each candidate in your pipeline contributed to it, we can share an overview of how predictions generate.
Offer forecast uses an ensemble model, which means that it’s composed of multiple smaller models that work together to make the best forecast. During the model training process, each smaller, specialist model becomes very good at recognizing specific types of patterns. Offer forecast then looks at cases where the specialist models gets it wrong, and adds more specialists to handle those more difficult cases. The forecast continues adding specialists that solve increasingly difficult problems until the combination of outputs from all the specialists is reasonably reliable. This way, offer forecast is able to recognize patterns, make data-driven decisions, and learn from new data, so it improves over time.
Machine learning can find complicated relationships that aren’t always easy to explain, understand, or untangle. However, by training offer forecast on huge amounts of data, we can trust the model’s decisions by comparing forecasts to actual hire dates.
How forecasts change over time
Offer forecast uses a number of factors in order to make a forecast.
Each of these factors is given a certain value and weight in the equation. When any of these factors change, it’s possible that the forecast might change too. However, each of these inputs might have a small effect on the overall forecast because so many different combinations of data are happening in the model. This means making one change might not shift your forecast, and it isn’t always easy to discern what change affected the forecast.