The prediction problem needs three inputs
To predict fit you need your measurements, the garment's finished measurements, and the ease built in. Retail currently supplies you with one, sometimes two, and never the third.
Why body shape is the wrong input
Body shape classification is not a fit variable. The dominant system in 3D body-shape research, FFIT, defines nine categories. Researchers analysing 1,679 3D body scans found it has never been scientifically verified, and that moving a tape measure by 1 cm reclassifies up to 40% of bodies.
A category that unstable cannot reliably predict anything.
What does determine fit
Direct comparison. Your shoulder width against the garment's shoulder-to-shoulder. Your waist against its finished waist. Numbers with defined meaning, not labels.
And crucially, ease. New Mexico State University Extension specifies 3 to 4 inches on the bust and 1 to 2 inches from nape to waist, and negative ease for stretch knit. Those determine how a garment feels, and they are rarely published.
What machine learning can and cannot do
On a public five-class body shape dataset, a VGG16 model only reliably separated two of the five categories. VGG19 failed entirely, and ResNet18 confused inverted triangle with hourglass.
That is classification failing on a task with no validated ground truth. Predicting garment fit from body shape is a harder problem built on that same unstable foundation.
Where prediction genuinely works
Compare two garments you already own to establish your own fit profile, then predict new purchases from that. This is reliable because it is calibrated to your body.
- Record how a brand's medium fits you, in numbers.
- Do the same for another size.
- Infer the pattern and apply it to future purchases.
This is a personal model with real evidence behind it, built entirely from your own data.
The ceiling
Predictions improve as garment data improves. Until brands publish finished measurements, any fit prediction remains an estimate built on incomplete information.
Related: what brands should publish and the body shape problem.