Two Regression Heads With a Negative, Heavy-Tailed Target
Problem statement
How would you train a shared neural network to predict dispute losses and processing-fee revenue? Losses are non-positive, mostly zero and heavy-tailed, while revenue is non-negative and much smaller in scale. Finance needs loss predictions in dollars that add up reliably for its forecasts.
Explain target transformations, loss balancing, monitoring and conversion back to dollars.
Show hint
Look at both the scale and the shape of each target's distribution, and at which data the normalization statistics may come from.
Relevant roles: ML Engineer, Data Scientist and Applied Scientist.