Delphine
ML Demand Forecasting for Footwear Retail
Multi-modal sell-through prediction for footwear: a photo of the shoe plus its attributes, and a buy/pass verdict before the season.
- Training rows
- 134,690
- Out-of-sample AUC
- 0.723
- Model features
- 105
The problem
Footwear buys are decided months before launch, looking at a photo and a spec sheet. Getting it wrong on the high side means clearing at negative margin; on the low side, missing the season.
How it works
A late-fusion neural network combines product photography (EfficientNet-B0) with categorical attributes — brand, model, material, color, gender, season — and numeric variables like price, and predicts weekly sell-through. The buyer works through a Telegram bot: send the photo, get the verdict with a calibrated uncertainty band.
Inside the system.
My perspective
Delphine belongs to the AI portfolio we build for Yáneken. It brings an engineering approach to a commercial decision: choosing footwear months before customers can tell us what they want. The buyer brings judgment; the system contributes a view of risk.
The design decision
Combine the product image with structured attributes and price. The published model uses a late-fusion network with EfficientNet-B0 for photography. Its output reaches the buyer through Telegram, together with an uncertainty band. The final buying decision stays with the person.
What the result shows
The published record contains 134,690 training rows, 105 features and an out-of-sample AUC of 0.723. AUC measures ranking discrimination across the evaluation set. It should not be read as 72.3% accuracy or a guaranteed result for an individual shoe.
How to read the evidence
The public record does not include the held-out sample definition or a comparison against a simple baseline. Those details are needed to assess generalization more fully. No reduction in markdowns or increase in buying margin is claimed here; the reported result is the model metric, with the buyer retaining the decision.
The path of a decision
- 01
The buyer provides a product photo through the Telegram interface.
- 02
The model combines visual information with attributes such as brand, material, season and price.
- 03
A buy/pass verdict and uncertainty band inform the buyer’s assessment before the season.
The thinking behind the work, in my essays:
What it does
- 01
Genuinely multi-modal
The product photo is a model input, not decoration: vision plus metadata in a single network.
- 02
Verdict with uncertainty
Buy or pass, with a calibrated uncertainty band — the model also says how confident it is.
- 03
In the buyer's flow
Telegram interface: the verdict arrives where the buyer already works, with no new system to learn.
Limits
0.723 out-of-sample AUC: the model ranks risk well, but does not replace the buyer's judgment on any individual decision. That is why the verdict ships with an uncertainty band.
Related projects
- IRISMultimodal AI for Retail Store Operations
- MeridianBayesian Marketing Mix Modeling for Retail
- AndreaAI Consumer Simulation for Marketing
Status and figures: August 2026
If you recognize this problem in your business, let’s talk. For an implementation with my team, the advisory work lives at Menlo & Oak.