@Anonymous wrote:
@CassieCard What do you think would be a good threshold to change scorecards between aged and non-aged, considering all your knowledge and experience in AI and predictive analytics, knowing everything that you know? You were giving an example of taking two years worth of data and then regressing and potentially determining scorecards from that? So can you give me more insight from your perspective?
My experience comes from translating the mathematical formulas to high performance code, so that millions of data points can be updated in near real-time. It's not a simple handoff from a mathematician/data scientist so I naturally learned a lot about the setting of bins from a large dataset.
A person could know everything about all the different systems used in predictive analytics, but without the data (payment behavior over time here) they would be speculating just like everyone else.
That said, it looks to me like 2 years of payment behavior seems good enough to make decent predictions, since that's the timeframe most talked about with trended data and credit scoring. If you read a lot about binning (scorecard segmentation here), you'll come across the number 10 used a lot as a general guideline for how many bins are usually needed, but 8 could be just as good - it depends on the dataset.
In that interview with Tom Quinn from FICO that we both listened to, he mentioned 14 scorecards. I immediately thought of 10 clean and 4 dirty. But I think that's just for FICO 8 models, since that's the score most people care about- the other models could each have different segmentation. There's also a possiblity of sub-binning as well - segmentation inside of segmentation, like the 3mo and 6mo thresholds.
My own 2 years of history isn't a very good predictor of where I'll be at in 10 years, but it's probably very good (never absolutely conclusive of course) at predicting how I'll end up 2 years from now. I highly doubt that 1 simple age metric - like AoOA - is used alone as a scorecard segmenter. It's probably used in combination with several factors, including but not limited to values such as Age of Revolving Credit History and number of accounts on file.
FICO now has decades of payment behavior from millions of people to analyze. At some point there will be diminishing returns, where practically everyone with say 20 years of credit history mostly ends up the same 10 years down the road. So I would expect to see scorecard segmentation every 2 to 4 years. Some might be longer than others, but this all depends on what regression over the data says - not some arbitrary decision made by a human. This is where WOE (weight of evidence) / IV (information value) comes in. Search for WOE/IV used in credit scoring and you'll see how these values are generated from a large sample dataset of payment behavior.
In the original dataset that FICO used, there was a large group of people that had similar data to my current profile. My score is a prediction of how I might end up a few years down the road based on what I share in common with people a few years ahead of me.
