New curve fitting algorithm for Rationalization

We’ve upgraded the Rationalization model with a new curve fitting approach to improve accuracy and reduce manual tuning.

What changed

Curve tournament:
the model now tests multiple diminishing-return curve families (e.g., exponential, power, hyperbolic, logistic variants) and automatically selects the best fit using AIC.
Non-linear least squares fitting:
instead of forcing the curve through the first/last points, we fit the full distribution to reduce over/under-estimation.
Smarter extension & contraction behavior:
improved tail handling without the “trimming” artifact, with a hard stop on extreme extensions.
No more width-depth slider:
assortment width is now driven by a minimum ROS threshold (more consistent and less manual).
Optional target choice count:
users can force an exact number of choices when needed.

What you’ll notice

More stable recommendations across clusters/seasons
Better alignment between modeled curves and planned “actual” choices (actual choices are assimilated and locked, then the remainder is rescaled to hit the productivity target)
“Calculate all” CTA that triggers the curve fitting API for all partitions in the list.

Read more about Rationalization in this help article.